CN116167798A - Data processing method, computer equipment and readable storage medium - Google Patents
Data processing method, computer equipment and readable storage medium Download PDFInfo
- Publication number
- CN116167798A CN116167798A CN202111385324.8A CN202111385324A CN116167798A CN 116167798 A CN116167798 A CN 116167798A CN 202111385324 A CN202111385324 A CN 202111385324A CN 116167798 A CN116167798 A CN 116167798A
- Authority
- CN
- China
- Prior art keywords
- feature
- target
- service
- conversion
- layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0241—Advertisements
- G06Q30/0242—Determining effectiveness of advertisements
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
- G06F16/43—Querying
- G06F16/435—Filtering based on additional data, e.g. user or group profiles
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0241—Advertisements
- G06Q30/0247—Calculate past, present or future revenues
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0241—Advertisements
- G06Q30/0251—Targeted advertisements
- G06Q30/0254—Targeted advertisements based on statistics
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Accounting & Taxation (AREA)
- Development Economics (AREA)
- Finance (AREA)
- Strategic Management (AREA)
- General Physics & Mathematics (AREA)
- Game Theory and Decision Science (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Economics (AREA)
- Entrepreneurship & Innovation (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Probability & Statistics with Applications (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Databases & Information Systems (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The embodiment of the application provides a data processing method, computer equipment and a readable storage medium, wherein the method can be applied to various scenes such as cloud technology, artificial intelligence, intelligent transportation, multimedia and the like, and comprises the following steps: acquiring target business data associated with a target object, and determining matching characteristics between the target object and the target business data; determining a target service triggering characteristic and a target service conversion characteristic; performing splicing processing on the matching features and the target service triggering features to obtain initial splicing triggering features, and determining triggering probability pre-estimation values of the target objects for the target service data based on the initial splicing triggering features; and performing splicing processing on the matching characteristic and the target service conversion characteristic to obtain an initial splicing conversion characteristic, and determining a conversion probability estimated value of the target object for the target service data based on the initial splicing conversion characteristic. The method and the device can improve the accuracy of the trigger probability predicted value and the conversion probability predicted value of the predicted target business data.
Description
Technical Field
The present disclosure relates to the field of internet technologies, and in particular, to a data processing method, a computer device, and a readable storage medium.
Background
With the development of multimedia technology, multimedia data (e.g., graphic data) has become a main carrier for acquiring information and enjoying entertainment in daily life. The service data recommendation system may deliver service data to the application client while distributing the teletext data to the application client, it being understood that the service data recommendation system typically determines service data for delivery to the application client based on click-through rate and conversion rate.
The existing service data recommendation system respectively trains a click rate estimation model and a conversion rate estimation model by using a deep learning model, and further respectively extracts click characteristics and conversion characteristics of service data (for example, service data G) through the click rate estimation model and the conversion rate estimation model. The conversion rate estimation model is obtained by training sample service data from click behavior to conversion behavior, namely, the sample service data have click behavior. However, the service data G does not have clicking behaviors when the conversion rate estimation model is used for prediction, and the conversion characteristics of the service data G extracted by the conversion rate estimation model are inaccurate due to inconsistent distribution of training data and prediction data of the conversion rate estimation model, so that the prediction accuracy of the conversion rate estimation model is reduced.
Disclosure of Invention
The embodiment of the application provides a data processing method, computer equipment and a readable storage medium, which can improve the accuracy of a trigger probability estimated value and a conversion probability estimated value of predicted target service data.
In one aspect, an embodiment of the present application provides a data processing method, including:
acquiring target business data associated with a target object, and determining matching characteristics between the target object and the target business data;
determining target service triggering characteristics and target service conversion characteristics of a target object aiming at target service data; the target service conversion feature is obtained by carrying out feature sharing transfer processing on the target service triggering feature;
performing splicing processing on the matching features and the target service triggering features to obtain initial splicing triggering features of the target object aiming at the target service data, and determining triggering probability pre-estimation values of the target object aiming at the target service data based on the initial splicing triggering features;
and performing splicing processing on the matching characteristic and the target service conversion characteristic to obtain an initial splicing conversion characteristic of the target object aiming at the target service data, and determining a conversion probability estimated value of the target object aiming at the target service data based on the initial splicing conversion characteristic.
An aspect of an embodiment of the present application provides a data processing apparatus, including:
the first determining module is used for acquiring target business data associated with the target object and determining matching characteristics between the target object and the target business data;
the second determining module is used for determining target service triggering characteristics and target service conversion characteristics of the target object aiming at the target service data; the target service conversion feature is obtained by carrying out feature sharing transfer processing on the target service triggering feature;
the first estimating module is used for carrying out splicing processing on the matching characteristic and the target service triggering characteristic to obtain an initial splicing triggering characteristic of the target object aiming at the target service data, and determining a triggering probability estimated value of the target object aiming at the target service data based on the initial splicing triggering characteristic;
the second estimating module is used for carrying out splicing processing on the matching characteristic and the target service conversion characteristic to obtain an initial splicing conversion characteristic of the target object aiming at the target service data, and determining a conversion probability estimated value of the target object aiming at the target service data based on the initial splicing conversion characteristic.
Wherein the first determining module comprises:
the data acquisition unit is used for acquiring a service data set associated with the target object and acquiring target service data from the service data set;
The feature acquisition unit is used for acquiring the pre-training object features associated with the target object and the pre-training service features associated with the service data sequence; the business data sequence is associated with the target object;
the feature embedding unit is used for inputting the object attribute of the target object, the service attribute of the target service data, the pre-training object feature and the pre-training service feature into the matching sub-network in the probability estimating network model; the matching sub-network comprises a matching embedding layer;
the feature embedding unit is used for carrying out feature embedding on the object attribute and the service attribute through the matching embedding layer to obtain an object feature corresponding to the object attribute and a service feature corresponding to the service attribute;
the vector determining unit is used for determining an object feature vector associated with the target object according to the object feature and the pre-training object feature in the matching sub-network, and determining a service feature vector associated with the target service data according to the service feature and the pre-training service feature;
and the dot product operation unit is used for carrying out dot product operation on the object feature vector and the service feature vector to obtain the matching feature between the target object and the target service data.
The service data sequence comprises a service trigger sequence and a service conversion sequence; the service trigger sequence comprises service data of the target object with trigger behavior, and the service conversion sequence comprises service data of the target object with conversion behavior;
The feature acquisition unit is specifically used for inputting the object attribute of the target object, the service trigger sequence associated with the target object and the service conversion sequence associated with the target object into the feature recognition network model; the feature recognition network model comprises a pre-training embedded layer, an attention processing layer and a pre-training connecting layer;
the feature acquisition unit is specifically configured to perform feature embedding on the object attribute, the service data in the service trigger sequence, and the service data in the service conversion sequence through the pre-training embedding layer, so as to obtain an object feature corresponding to the object attribute, a trigger service feature corresponding to the service data in the service trigger sequence, and a conversion service feature corresponding to the service data in the service conversion sequence;
the feature acquisition unit is specifically configured to input an object feature, a trigger service feature and a conversion service feature to the attention processing layer, and perform attention processing on the object feature, the trigger service feature and the conversion service feature through the attention processing layer to obtain an attention object feature corresponding to the object feature, an attention trigger feature corresponding to the trigger service feature and an attention conversion feature corresponding to the conversion service feature;
The feature acquisition unit is specifically configured to input an object feature, an attention triggering feature and an attention conversion feature to the pre-training connection layer, perform feature fusion on the object feature and the attention object feature through the pre-training connection layer, obtain a pre-training object feature associated with the target object, and generate, in the pre-training connection layer, a pre-training service feature corresponding to service data included in the service triggering sequence and the service conversion sequence according to the attention triggering feature and the attention conversion feature.
The matching sub-network comprises an object full-connection layer, a service full-connection layer and a characteristic average layer; the service data sequence comprises one or more service data, and each service data corresponds to a pre-training service feature respectively;
the vector determining unit is specifically configured to input the object feature to the object full-connection layer in the matching sub-network, and perform full-connection processing on the object feature through the object full-connection layer to obtain a full-connection object feature;
the vector determining unit is specifically used for carrying out feature fusion on the fully-connected object features and the pre-trained object features to obtain an object feature vector associated with the target object;
The vector determining unit is specifically configured to input service features to a service full-connection layer, and perform full-connection processing on the service features through the service full-connection layer to obtain full-connection service features;
the vector determining unit is specifically configured to input one or more pre-training service features to the feature averaging layer, and average the one or more pre-training service features through the feature averaging layer to obtain average pre-training service features;
the vector determining unit is specifically configured to perform feature fusion on the full-connection service feature and the average pre-training service feature, so as to obtain a service feature vector associated with the target service data.
Wherein the probability estimation network model further comprises a sequencing subnetwork; the sequencing sub-network comprises an input network layer, a parameter sharing network layer and a multi-layer perception network layer;
the second determination module includes:
the characteristic determining unit is used for determining the sharing attribute characteristic of the target object aiming at the target service data through the input network layer;
the feature generation unit is used for inputting the shared attribute feature into the parameter sharing network layer, and generating an initial service triggering feature and an initial service conversion feature of the target object aiming at the target service data through the parameter sharing network layer;
The feature transfer unit is used for inputting the initial service triggering feature and the initial service conversion feature into a multi-layer sensing network layer, performing full-connection processing on the initial service triggering feature through the multi-layer sensing network layer to obtain the target service triggering feature of the target object aiming at the target service data, and performing feature sharing transfer processing on the target service triggering feature and the initial service conversion feature in the multi-layer sensing network layer to obtain the target service conversion feature of the target object aiming at the target service data.
The input network layer comprises a characteristic embedding layer and a characteristic splicing layer; the service data sequence comprises one or more service data, and each service data corresponds to a pre-training service feature respectively;
the feature determining unit is specifically configured to input an object attribute of a target object, a service attribute of target service data, and a context attribute associated with the target object to the feature embedding layer, and perform feature embedding on the object attribute, the service attribute, and the context attribute through the feature embedding layer to obtain an object feature corresponding to the object attribute, a service feature corresponding to the service attribute, and a context feature corresponding to the context attribute;
The feature determining unit is specifically configured to perform average processing on one or more pre-training service features to obtain average pre-training service features;
the feature determining unit is specifically configured to input an object feature, a pre-training object feature, a service feature and an average pre-training service feature into the feature stitching layer, perform feature stitching on the object feature and the pre-training object feature in the feature stitching layer to obtain a stitched object feature, and perform feature stitching on the service feature and the average pre-training service feature to obtain a stitched service feature;
the feature determining unit is specifically configured to perform feature stitching on the stitching object feature, the stitching service feature, and the context feature, so as to obtain a sharing attribute feature of the target object for the target service data.
The parameter sharing network layer comprises a sharing full-connection layer, a weight learning layer, a feature classification layer, a first triggering full-connection layer, a first shallow conversion full-connection layer and a first deep conversion full-connection layer; the initial business transformation features comprise initial shallow transformation features and initial deep transformation features;
the feature generation unit is specifically configured to input the shared attribute feature to a shared full-connection layer, and perform full-connection processing on the shared attribute feature through the shared full-connection layer to obtain a full-connection shared feature;
The feature generation unit is specifically configured to input the full-connection shared feature to the weight learning layer, and perform feature weighting processing on the full-connection shared feature through the weight learning layer to obtain a weight shared feature;
the feature generation unit is specifically used for inputting the weight sharing features into the feature classification layer, and carrying out feature classification on the weight sharing features through the feature classification layer to obtain a trigger distribution vector, a shallow conversion distribution vector and a deep conversion distribution vector of a target object aiming at target service data;
the feature generation unit is specifically configured to input a trigger distribution vector to a first trigger full-connection layer, perform full-connection processing on the trigger distribution vector through the first trigger full-connection layer, and generate an initial service trigger feature of a target object for target service data;
the feature generation unit is specifically configured to input a shallow conversion distribution vector to a first shallow conversion full-connection layer, perform full-connection processing on the shallow conversion distribution vector through the first shallow conversion full-connection layer, and generate an initial shallow conversion feature of the target object for the target service data;
the feature generation unit is specifically configured to input a deep conversion distribution vector to a first deep conversion full-connection layer, and perform full-connection processing on the deep conversion distribution vector through the first deep conversion full-connection layer, so as to generate an initial deep conversion feature of the target object for the target service data.
The weight learning layer comprises weight learning components corresponding to object features, business features, context features, pre-training object features and average pre-training business features respectively; the feature classification layer comprises a triggering feature classification layer, a shallow conversion feature classification layer and a deep conversion feature classification layer;
the feature generation unit is specifically configured to input the full-connection sharing feature to each weight learning component, and perform feature weighting processing on the full-connection sharing feature through each weight learning component to obtain a weight sharing feature corresponding to each weight learning component;
the feature generation unit is specifically used for carrying out feature fusion on the weight sharing features corresponding to each weight learning component respectively to obtain fusion sharing features;
the feature generation unit is specifically configured to input the fusion shared feature to the trigger feature classification layer, and perform feature classification on the fusion shared feature through the trigger feature classification layer to obtain a trigger distribution vector of the target object for the target service data;
the feature generation unit is specifically configured to input the fusion shared feature to a shallow conversion feature classification layer, and perform feature classification on the fusion shared feature through the shallow conversion feature classification layer to obtain a shallow conversion distribution vector of the target object for the target service data;
The feature generation unit is specifically configured to input the fusion shared feature to a deep conversion feature classification layer, and perform feature classification on the fusion shared feature through the deep conversion feature classification layer to obtain a deep conversion distribution vector of the target object for the target service data.
The multi-layer sensing network layer comprises a second triggering full-connection layer, a second shallow conversion full-connection layer, a shallow connection layer, a second deep conversion full-connection layer and a deep connection layer; the initial business transformation features comprise initial shallow transformation features and initial deep transformation features; the target business transformation features comprise target shallow transformation features and target deep transformation features;
the feature transfer unit is specifically configured to input an initial service triggering feature to the second triggering full-connection layer, and perform full-connection processing on the initial service triggering feature through the second triggering full-connection layer to obtain a target service triggering feature of the target object for target service data;
the feature transfer unit is specifically configured to input the target service triggering feature and the initial shallow conversion feature to a shallow connection layer, and perform feature splicing on the target service triggering feature and the initial shallow conversion feature through the shallow connection layer to obtain a shared shallow splicing feature;
The feature transfer unit is specifically configured to input the shared shallow splicing feature to a second shallow conversion full-connection layer, and perform full-connection processing on the shared shallow splicing feature through the second shallow conversion full-connection layer to obtain a target shallow conversion feature of the target object for the target service data;
the feature transfer unit is specifically used for inputting the target shallow layer transformation feature and the initial deep layer transformation feature into the deep connecting layer, and performing feature splicing on the target shallow layer transformation feature and the initial deep layer transformation feature through the deep connecting layer to obtain a shared deep splicing feature;
the feature transfer unit is specifically configured to input the shared deep splicing feature to the second deep conversion full-connection layer, and perform full-connection processing on the shared deep splicing feature through the second deep conversion full-connection layer, so as to obtain a target deep conversion feature of the target object for the target service data.
The ordering sub-network further comprises a characteristic connecting layer and a dot product full connecting layer;
the first estimating module comprises:
the first splicing unit is used for inputting the matching characteristic and the target service triggering characteristic into the characteristic connecting layer, and splicing the matching characteristic and the target service triggering characteristic through the characteristic connecting layer to obtain an initial splicing triggering characteristic of the target object aiming at the target service data;
The first processing unit is used for inputting the initial splicing triggering characteristics into the dot product full-connection layer, and carrying out full-connection processing on the initial splicing triggering characteristics through the dot product full-connection layer to obtain target splicing triggering characteristics;
the first estimating unit is used for determining a trigger probability estimated value of the target object aiming at the target service data according to the target splicing trigger characteristics.
The ordering sub-network further comprises a characteristic connecting layer and a dot product full connecting layer; the conversion probability estimated value comprises a shallow conversion probability estimated value and a deep conversion probability estimated value; the target business transformation features comprise target shallow transformation features and target deep transformation features;
the second estimating module comprises:
the second splicing unit is used for inputting the matching features and the target shallow layer conversion features into the feature connecting layer, and splicing the matching features and the target shallow layer conversion features through the feature connecting layer to obtain initial spliced shallow layer conversion features of the target object aiming at the target service data;
the second processing unit is used for inputting the initial splicing shallow layer conversion characteristics to the dot product full-connection layer, and carrying out full-connection processing on the initial splicing shallow layer conversion characteristics through the dot product full-connection layer to obtain target splicing shallow layer conversion characteristics;
The second estimating unit is used for determining shallow conversion probability estimated values of the target object aiming at the target service data according to the target splicing shallow conversion characteristics;
the third splicing unit is used for inputting the matching feature and the target deep conversion feature into the feature connecting layer, and splicing the matching feature and the target deep conversion feature through the feature connecting layer to obtain an initial spliced deep conversion feature of the target object aiming at the target service data;
the third processing unit is used for inputting the initial splicing deep conversion characteristics into the dot product full-connection layer, and carrying out full-connection processing on the initial splicing deep conversion characteristics through the dot product full-connection layer to obtain target splicing deep conversion characteristics;
and the third pre-estimating unit is used for determining a deep conversion probability estimated value of the target object aiming at the target service data according to the target splicing deep conversion characteristics.
In one aspect, an embodiment of the present application provides a data processing method, including:
acquiring sample service data associated with a sample object, and determining sample matching characteristics between the sample object and the sample service data through an initial probability pre-estimated network model;
in an initial probability pre-estimated network model, determining sample service triggering characteristics and sample service conversion characteristics of a sample object aiming at sample service data; the sample service conversion feature is obtained by carrying out feature sharing transfer processing on the sample service triggering feature;
Performing splicing processing on the sample matching characteristics and the sample service triggering characteristics to obtain sample splicing triggering characteristics of the sample object aiming at the sample service data, and determining a sample triggering probability estimated value of the sample object aiming at the sample service data based on the sample splicing triggering characteristics;
performing splicing processing on the sample matching characteristics and the sample service conversion characteristics to obtain sample splicing conversion characteristics of the sample object aiming at the sample service data, and determining a sample conversion probability estimated value of the sample object aiming at the sample service data based on the sample splicing conversion characteristics;
based on sample label information, sample triggering probability pre-estimation value and sample conversion probability pre-estimation value of a sample object aiming at sample service data, carrying out parameter adjustment on an initial probability pre-estimation network model, and taking the initial probability pre-estimation network model after parameter adjustment as a probability pre-estimation network model; the probability prediction network model is used for predicting a trigger probability predicted value and a conversion probability predicted value of the target object aiming at the target service data.
An aspect of an embodiment of the present application provides a data processing apparatus, including:
the first sample determining module is used for acquiring sample service data associated with the sample object and determining sample matching characteristics between the sample object and the sample service data through an initial probability pre-estimated network model;
The second sample determining module is used for determining sample service triggering characteristics and sample service conversion characteristics of the sample object aiming at the sample service data in the initial probability pre-estimated network model; the sample service conversion feature is obtained by carrying out feature sharing transfer processing on the sample service triggering feature;
the first sample prediction module is used for performing splicing processing on the sample matching characteristics and the sample service triggering characteristics to obtain sample splicing triggering characteristics of the sample object aiming at the sample service data, and determining a sample triggering probability estimated value of the sample object aiming at the sample service data based on the sample splicing triggering characteristics;
the second sample prediction module is used for performing splicing processing on the sample matching characteristics and the sample service conversion characteristics to obtain sample splicing conversion characteristics of the sample object aiming at the sample service data, and determining a sample conversion probability estimated value of the sample object aiming at the sample service data based on the sample splicing conversion characteristics;
the parameter adjustment module is used for carrying out parameter adjustment on the initial probability prediction network model based on sample label information, sample triggering probability prediction value and sample conversion probability prediction value of the sample object aiming at the sample service data, and taking the initial probability prediction network model after parameter adjustment as a probability prediction network model; the probability prediction network model is used for predicting a trigger probability predicted value and a conversion probability predicted value of the target object aiming at the target service data.
In one aspect, a computer device is provided, including: a processor and a memory;
the processor is connected to the memory, wherein the memory is configured to store a computer program, and when the computer program is executed by the processor, the computer device is caused to execute the method provided in the embodiment of the application.
In one aspect, the present application provides a computer readable storage medium storing a computer program adapted to be loaded and executed by a processor, so that a computer device having the processor performs the method provided in the embodiments of the present application.
In one aspect, the present application provides a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the method provided in the embodiments of the present application.
According to the method and the device for determining the target business triggering characteristics, after the target business data associated with the target object are acquired, the matching characteristics between the target object and the target business data can be determined, and then the target business triggering characteristics and the target business transformation characteristics of the target object aiming at the target business data are determined. The target service conversion feature is obtained by carrying out feature sharing transfer processing on the target service triggering feature. It can be appreciated that by performing the stitching process on the matching feature and the target service triggering feature, an initial stitching triggering feature of the target object for the target service data can be obtained, where the initial stitching triggering feature can be used to generate a triggering probability pre-estimated value of the target object for the service data. Similarly, by performing stitching processing on the matching feature and the target service conversion feature, an initial stitching conversion feature of the target object for the target service data can be obtained, and the initial stitching conversion feature can be used for generating a conversion probability estimated value of the target object for the service data. Based on the above, the feature sharing transfer processing is performed on the target service triggering feature, so that the target service triggering feature corresponding to the service triggering task can be transferred to the target service conversion feature corresponding to the service conversion task, which is equivalent to sharing the bottom layer parameter corresponding to the service triggering task to the service conversion task, thereby generating the target service conversion feature with higher accuracy. In addition, the embodiment of the application can integrate the matching characteristics between the target object and the target service data into the click probability predicted value and the conversion probability predicted value, fully utilize the correlation information between the target object and the target service data, simultaneously generate the trigger probability predicted value and the conversion probability predicted value, and improve the accuracy of the trigger probability predicted value and the conversion probability predicted value of the predicted target service data.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the related art, the drawings that are required to be used in the embodiments or the related technical descriptions will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and other drawings may be obtained according to the drawings without inventive effort for a person having ordinary skill in the art.
Fig. 1 is a schematic structural diagram of a network architecture according to an embodiment of the present application;
fig. 2 is a schematic view of a scenario for data interaction according to an embodiment of the present application;
FIG. 3 is a schematic flow chart of a data processing method according to an embodiment of the present application;
FIG. 4 is a schematic flow chart of a data processing method according to an embodiment of the present application;
FIG. 5 is a schematic structural diagram of a feature recognition network model according to an embodiment of the present application;
FIG. 6 is a schematic flow chart of a data processing method according to an embodiment of the present application;
FIG. 7 is a schematic flow chart of a data processing method according to an embodiment of the present application;
FIG. 8a is a schematic diagram of a probabilistic predictive network model according to an embodiment of the present application;
FIG. 8b is a schematic structural diagram of a probability estimation network model according to an embodiment of the present disclosure;
fig. 9 is a schematic flow chart of service data recommendation provided in an embodiment of the present application;
FIG. 10 is a schematic flow chart of a data processing method according to an embodiment of the present application;
FIG. 11 is a schematic diagram of a data processing apparatus according to an embodiment of the present application;
FIG. 12 is a schematic diagram of a data processing apparatus according to an embodiment of the present disclosure;
fig. 13 is a schematic structural diagram of a computer device according to an embodiment of the present application.
Detailed Description
The following description of the embodiments of the present application will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are only some, but not all, of the embodiments of the present application. All other embodiments, which can be made by one of ordinary skill in the art based on the embodiments herein without making any inventive effort, are intended to be within the scope of the present application.
It should be appreciated that artificial intelligence (Artificial Intelligence, AI for short) is the intelligence of a person using a digital computer or a machine controlled by a digital computer to simulate, extend and extend the environment, sense the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is an integrated technology of computer science that attempts to understand the essence of intelligence and to produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence, i.e. research on design principles and implementation methods of various intelligent machines, enables the machines to have functions of sensing, reasoning and decision.
The artificial intelligence technology is a comprehensive subject, and relates to the technology with wide fields, namely the technology with a hardware level and the technology with a software level. Artificial intelligence infrastructure technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning, automatic driving, intelligent traffic and other directions.
The solution provided in the embodiments of the present application mainly relates to an artificial intelligence Machine Learning (ML) technology. Machine learning is a multi-domain interdisciplinary, involving multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, and the like. It is specially studied how a computer simulates or implements learning behavior of a human to acquire new knowledge or skills, and reorganizes existing knowledge structures to continuously improve own performance. Machine learning is the core of artificial intelligence, a fundamental approach to letting computers have intelligence, which is applied throughout various areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, confidence networks, reinforcement learning, transfer learning, induction learning, teaching learning, and the like.
Specifically, referring to fig. 1, fig. 1 is a schematic structural diagram of a network architecture according to an embodiment of the present application. As shown in fig. 1, the network architecture may include a service server 2000 and a cluster of terminal devices. Wherein the cluster of terminal devices may in particular comprise one or more terminal devices, the number of terminal devices in the cluster of terminal devices will not be limited here. As shown in fig. 1, the plurality of terminal devices may specifically include a terminal device 3000a, a terminal device 3000b, terminal devices 3000c, …, a terminal device 3000n; the terminal devices 3000a, 3000b, 3000c, …, 3000n may be directly or indirectly connected to the service server 2000 through a wired or wireless communication manner, so that each terminal device may interact with the service server 2000 through the network connection.
Wherein each terminal device in the terminal device cluster may include: intelligent terminals with data processing functions such as intelligent televisions, intelligent mobile phones, tablet computers, notebook computers, desktop computers, intelligent home, wearable equipment and vehicle-mounted terminals. It should be understood that each terminal device in the terminal device cluster shown in fig. 1 may be integrally provided with an application client, and when the application client runs in each terminal device, the application client may perform data interaction with the service server 2000 shown in fig. 1. The application client may be an independent client or an embedded sub-client integrated in a certain client, which is not limited in this application.
The application client may specifically include a browser, a vehicle-mounted client, a smart home client, an entertainment client, a multimedia client (e.g., a video client), a social client, an information client (e.g., a news client), and other clients with data processing functions. The vehicle-mounted terminal can be an intelligent terminal in an intelligent traffic scene, and the application client on the vehicle-mounted terminal can be the vehicle-mounted client.
The service server 2000 may be a server corresponding to an application client, where the service server 2000 may be an independent physical server, may be a server cluster or a distributed system formed by a plurality of physical servers, or may be a cloud server that provides cloud services, a cloud database, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDNs, and basic cloud computing services such as big data and an artificial intelligence platform.
For ease of understanding, the embodiment of the present application may select one terminal device from the plurality of terminal devices shown in fig. 1 as the target terminal device. For example, the embodiment of the present application may use the terminal device 3000b shown in fig. 1 as a target terminal device, and an application client having a data processing function may be integrated in the target terminal device. At this time, the target terminal device may perform data interaction between the application client and the service server 2000.
For easy understanding, in the embodiment of the present application, the user corresponding to the target terminal device may be collectively referred to as a target object, and in the embodiment of the present application, service data recommended to the target object by the service server 2000 may be collectively referred to as recommended service data, and in the embodiment of the present application, service data sent by an advertiser (i.e. a person who is a sales promotion or provides a service, who is a legal person who is self or entrusted with others to design, make, and issue an advertisement, other economic organizations, or individuals) to the service server 2000 may be collectively referred to as service data to be recommended. The service data in the embodiment of the present application may be advertisement data (i.e. advertisement).
It will be appreciated that the target object may send a service data acquisition request to the service server 2000 through an application client in the target terminal device. In this way, after receiving the service data acquisition request, the service server 2000 may acquire a service data set associated with the target object, acquire recommended service data for recommending to the target object from the service data set, and then send the recommended service data to the application client corresponding to the target object. Accordingly, the application client may display the recommended service data in the application client after receiving the recommended service data issued by the service server 2000, so that the target object determines whether to perform a triggering operation for the recommended service data.
The advertisement exposure may represent that the target object observes the exposed product (i.e., recommended service data) in the application client, and the triggering operation performed by the target object for the recommended service data may be three types. The first category is advertisement clicking, and the advertisement clicking can represent clicking actions of a target object on recommended service data; the second type is advertisement shallow conversion, wherein the advertisement shallow conversion can represent shallow conversion behavior of recommended service data after a target object clicks, and the shallow conversion behavior can be downloading, activating, registering and the like; the third type is advertisement deep conversion, and the advertisement deep conversion can represent deep conversion behavior of recommended service data after a target object clicks, wherein the deep conversion behavior can be pay, retention in the next day and the like.
The shallow conversion behavior is based on the clicking behavior, namely the clicking behavior is followed by the shallow conversion behavior; the deep transformation behavior is based on the shallow transformation behavior, i.e. the existing shallow transformation behavior has the deep transformation behavior.
It should be understood that, in the embodiment of the present application, the behaviors corresponding to the above three types of triggering operations may be measured by different indicators. The ratio from advertisement exposure to advertisement click, i.e., the click-through amount of the advertisement divided by the display amount, can be expressed by CTR (Click Through Rate, click-through rate, simply click-through rate); the ratio from advertisement click-through to advertisement Conversion, i.e., the Conversion of an advertisement divided by the click-through, can be expressed by CVR (Conversion Rate). Wherein the CVR can be divided into CVR 1 (i.e., shallow conversion) and CVR 2 (i.e., deep conversion) by CVR 1 The ratio from advertisement click to advertisement shallow conversion, i.e. advertisement shallow conversion divided by click through, can be expressed by CVR 2 The ratio from advertisement click-through to advertisement deep conversion, i.e., the amount of deep conversion of an advertisement divided by the click-through, may be expressed.
It should be understood that, when the advertiser sends the service data to be recommended, the service target corresponding to the service data to be recommended needs to be set, where the service target may be divided into a shallow target advertisement and a deep target advertisement. An ad serving target refers to a bid target of the ad including an ad serving target bid, i.e., a bid pattern of a single bid for an ad serving indicator, where the ad serving target bid may represent a bid for an ad serving conversion behavior (i.e., ad serving indicator, e.g., download, activate, register), the ad serving target ad requiring an estimated click-through rate and ad serving conversion rate; deep targeted advertising refers to a bidding target of advertising that includes a shallow targeted bid and a deep targeted bid, i.e., a bidding pattern of re-bidding on a single bid for a deep bid, where deep targeted bid may represent a bid for deep conversion activity (i.e., deep bid, e.g., paid, reserved next day), deep targeted advertising requires estimated click-through, shallow conversion, and deep conversion. In addition, embodiments of the present application may refer to shallow targeted advertisements as single targeted advertisements and deep targeted advertisements as multi-targeted advertisements.
The bayesian network in the embodiment of the application is a probability graph model, which is also called a directed acyclic graph model. The Bayesian network is a graph network model for describing uncertainty causal relation among variables, and consists of nodes, directed links and a node probability table, wherein the directed links represent causal dependency relation among the nodes.
It may be appreciated that the above network framework may be applicable to an advertisement recommendation field, where a service scenario of the advertisement recommendation field may specifically include: advertisement distribution scenes, advertisement search scenes, advertisement viewing scenes, and the like, and specific business scenes will not be listed here one by one. It should be understood that the present application is not limited to the presentation of business data.
For example, in an advertisement distribution scenario, a computer device (e.g., the business server 2000 described above) may analyze a user representation of a target object to determine recommended business data (e.g., business data G) that may be of interest to the target object 1 ) Further, when the target object opens the application client, the content and the service data G are distributed 1 Pushing to the application client. The distribution content may be video data, graphic data, audio data, or the like, which is not limited herein.
For another example, in an advertisement search scenario, a computer device (e.g., the business server 2000 described above) may analyze search text entered by a target object in an application client to determine recommended business data (e.g., business data G) that may be of interest to the target object 2 ) Further, the search content and the service data G corresponding to the search text are searched 2 Pushing to the application client. The search content may be video data, graphic data, audio data, or the like, which is not limited herein.
For another example, in an advertisement viewing scenario, a computer device (e.g., the business server 2000 described above) may provide a user interface toAnalysis of the context of the viewing of the target object in the application client determines recommended business data (e.g., business data G) that may be of interest to the target object 3 ) And then the service data G 3 Pushing to the application client.
It will be appreciated that in the specific embodiments of the present application, related data relating to object properties, context properties, etc. are required to obtain user permissions or agreements when the above embodiments of the present application are applied to specific products or technologies, and the collection, use and processing of related data is required to comply with relevant laws and regulations and standards of the relevant countries and regions.
For ease of understanding, further, please refer to fig. 2, fig. 2 is a schematic diagram of a scenario for data interaction according to an embodiment of the present application. The server 20a shown in fig. 2 may be the service server 2000 in the embodiment corresponding to fig. 1, and the terminal device 20b shown in fig. 2 may be the target terminal device in the embodiment corresponding to fig. 1. Wherein an application client is installed on the terminal device 20b, and the application client can be used for displaying recommended service data associated with the object 20c corresponding to the terminal device 20 b.
It can be understood that the object 20c may send a service data acquisition request to the server 20a through an application client in the terminal device 20b, so that the application client may view multimedia data (e.g. video data, graphic data) attached with its own interest and view recommended service data matched with itself. For example, when the object 20c performs a search operation for an application client, the application client may send a service data acquisition request to the server 20a to cause the server 20a to return recommended service data associated with the search content while returning the search content. For another example, when the object 20c performs an open operation for an application client, the application client may send a service data acquisition request to the server 20a to cause the server 20a to return recommended service data associated with the distribution content while returning the distribution content. For another example, when the object 20c performs an information flow viewing operation with respect to an application client, the application client may send a service data acquisition request to the server 20a to cause the server 20a to return recommended service data associated with the information flow content while returning the information flow content.
For easy understanding, in this embodiment of the present application, the specific processes of the opening operation, the information flow viewing operation, and the like are described by taking the example that the service data acquisition request is triggered by the searching operation, and the description of the searching operation will be omitted here.
As shown in fig. 2, the server 20a may receive a service data acquisition request sent by an application client, and further acquire a service data set (e.g., service data set 21 b) associated with the object 20c in the service database 21 a. The service database 21a may include a plurality of databases, and the plurality of databases may include a database 22a, a database 22b, …, a database 22n, and the databases 22a, 22b, …, and 22n may be used to store different types of service data. For example, database 22a may be used to store business data associated with animals, database 22b may be used to store business data associated with plants, …, database 22n may be used to store business data associated with space.
Where the search text of the object 20c is cat, the server 20a may obtain the service data set 21b associated with the object 20c from the database 22 a. It will be appreciated that the server 20a may construct an object vector of the object 20c and a service vector of the service data in the database 22a, and further determine a vector similarity between the object vector and the service vector by using a cosine similarity or euclidean distance or the like, so as to obtain service data of which the vector similarity satisfies a vector threshold, and construct the service data obtained in the database 22a into the service data set 21b. The service data set 21b may include a plurality of service data, and the plurality of service data may include service data 23a, service data 23b, …, and service data 23m.
Alternatively, it should be understood that the server 20a may also obtain the service data for forming the service data set 21b from the databases 22a, 22b, …, and 22n at the same time, and this is illustrated by taking the service data 23a, the service data 23b, …, and the service data 23m obtained from the database 22a as an example.
As shown in fig. 2, after the service data set 21b is acquired, the server 20a may sequentially acquire target service data from the service data set 21b, and the acquired target service data is exemplified as service data 23 a. The server 20a may determine matching characteristics between the object 20c and the business data 23a, and thus determine target business trigger characteristics and target business transformation characteristics for the object 20c for the business data 23 a. The target service triggering characteristic corresponds to a characteristic of the object 20c aiming at the click behavior, the target service conversion characteristic corresponds to a characteristic of the object 20c aiming at the conversion behavior, and the target service conversion characteristic is obtained by carrying out characteristic sharing transfer processing on the target service triggering characteristic.
Further, the server 20a may perform a stitching process on the matching feature and the target service triggering feature to obtain an initial stitching triggering feature of the object 20c for the service data 23a, and further determine a triggering probability estimated value of the object 20c for the service data 23a based on the initial stitching triggering feature. Meanwhile, the server 20a may perform a stitching process on the matching feature and the target service transformation feature to obtain an initial stitching transformation feature of the object 20c for the service data 23a, and further determine a transformation probability estimated value of the object 20c for the service data 23a based on the initial stitching transformation feature. Wherein the trigger probability predicted value corresponds to the predicted value of the object 20c for the click behavior, and the conversion probability predicted value corresponds to the predicted value of the object 20c for the conversion behavior.
As shown in fig. 2, after determining the trigger probability predicted value and the conversion probability predicted value corresponding to each service data in the service data set 21b, the server 20a may sort each service data in the service data set 21b based on the trigger probability predicted value corresponding to each service data and the conversion probability predicted value corresponding to each service data, obtain recommended service data from each service data after sorting, and then use the recommended service data as service data matched with the object 20c, and return the recommended service data to the application client.
Therefore, after the matching characteristics between the target object and the target service data are determined, the matching characteristics are integrated into the target service triggering characteristics corresponding to the service triggering tasks, the triggering probability predicted value with higher accuracy is obtained, and the matching characteristics are integrated into the target service conversion characteristics corresponding to the service conversion tasks, so that the conversion probability predicted value with higher accuracy is obtained. In addition, by carrying out feature sharing transfer processing on the target service triggering feature, the accuracy of determining the target service conversion feature can be improved, and further the accuracy of predicting the triggering probability predicted value and the conversion probability predicted value of the target service data is improved based on the target service triggering feature and the target service conversion feature.
Further, referring to fig. 3, fig. 3 is a flow chart of a data processing method according to an embodiment of the present application. The method may be executed by a server, or may be executed by an application client, or may be executed by a server and an application client together, where the server may be the server 20a in the embodiment corresponding to fig. 2, and the application client may be the application client in the embodiment corresponding to fig. 2. For ease of understanding, embodiments of the present application will be described in terms of this method being performed by a server. The data processing method may include the following steps S101 to S104:
step S101, acquiring target business data associated with a target object, and determining matching characteristics between the target object and the target business data;
specifically, the server may obtain a set of business data associated with the target object, and obtain the target business data from the set of business data. Further, the server may obtain pre-trained object features associated with the target object, and pre-trained business features associated with the business data sequence. Wherein the sequence of business data is associated with the target object. Further, the server may input the object attributes of the target object, the traffic attributes of the target traffic data, the pre-trained object features, and the pre-trained traffic features to matching sub-networks in the probabilistic predictive network model. Wherein the matching sub-network comprises a matching embedding layer. Further, the server can perform feature embedding on the object attribute and the service attribute through the matching embedding layer to obtain an object feature (namely, a user side feature) corresponding to the object attribute and a service feature (namely, an advertisement side feature and an article feature) corresponding to the service attribute. Further, the server may determine, in the matching sub-network, an object feature vector associated with the target object based on the object feature and the pre-trained object feature, and determine a traffic feature vector associated with the target traffic data based on the traffic feature and the pre-trained traffic feature. Further, the server may perform a dot product operation on the object feature vector and the service feature vector to obtain a matching feature between the target object and the target service data.
It can be understood that the server can acquire the service data set associated with the target object through a recall model, and the recall model can find out service data with high correlation with the target object from massive service data as much as possible. The recall model may, among other things, comprehensively consider object attributes of the target object, business attributes of the business data, context attributes associated with the target object, and search terms associated with the target object. Wherein, the search word can represent the search text input by the target object when the search operation is performed in the application client.
The recall model can be divided into a first type recall model and a second type recall model. The first type recall model uses keyword-based hard matching, such as gender (male and female) in service data orientation, namely the service data is divided into two types according to the service data orientation, wherein one type is the service data corresponding to male and the other type is the service data corresponding to female; for another example, search terms (e.g., a cat) in the text are searched, i.e., business data associated with the cat is obtained. The second recall model uses soft matching based on vectors, namely, constructing an object vector of a target object and a service vector of service data, and further determining the vector similarity between the object vector and the service vector by using cosine similarity or Euclidean distance and other modes.
It will be appreciated that object attributes may be used to represent target objects and business attributes may be used to represent business data. Object attributes may include, but are not limited to, object identification of the target object, base attributes, and behavior attributes (i.e., behavior interests); the business attributes may include, but are not limited to, business identifications of business data (i.e., advertisement identifications), business data object identifications (i.e., advertiser identifications), business data categories (i.e., advertisement categories, e.g., application client advertisements), business data text (i.e., semantic features), and business data images (i.e., image features).
Wherein the object identifications of the target objects are unique, one target object corresponding to each object identification; the service identity of the service data is unique, one service data corresponding to each service identity. It should be appreciated that the target object need not have all of the object attributes and the business data need not have all of the business attributes.
The matching sub-network comprises an object full-connection layer, a service full-connection layer and a characteristic average layer; the service data sequence includes one or more service data, each corresponding to a respective one of the pre-trained service features. It should be appreciated that the specific process by which the server determines the object feature vector and the traffic feature vector in the matching sub-network may be described as: the server can input the object features into the object full-connection layer in the matching sub-network, and the object features are subjected to full-connection processing through the object full-connection layer to obtain full-connection object features. Further, the server may perform feature fusion on the fully connected object features and the pre-trained object features to obtain an object feature vector associated with the target object. Further, the server may input the service feature to the service full-connection layer, and perform full-connection processing on the service feature through the service full-connection layer to obtain a full-connection service feature. Further, the server may input one or more pre-training service features to the feature averaging layer, and average the one or more pre-training service features through the feature averaging layer to obtain an average pre-training service feature. Further, the server may perform feature fusion on the full-connection service feature and the average pre-training service feature to obtain a service feature vector associated with the target service data.
The probability estimation network model is obtained by performing iterative training on the initial probability estimation network model, and the specific process of performing iterative training on the initial probability estimation network model by the server to obtain the probability estimation network model can be described in the following embodiment corresponding to fig. 10 in step S301 to step S305.
It should be appreciated that the probabilistic predictive network model further comprises a ranking sub-network, wherein the ranking sub-network comprises an input network layer, a parameter sharing network layer and a multi-layer aware network layer, which may be used to perform step S102 described below.
Step S102, determining target service triggering characteristics and target service conversion characteristics of a target object aiming at target service data;
specifically, the server may determine, through the input network layer, a shared attribute characteristic of the target object for the target service data. Further, the server may input the shared attribute feature to a parameter sharing network layer, and generate an initial service trigger feature and an initial service conversion feature of the target object for the target service data through the parameter sharing network layer. Further, the server may input the initial service triggering feature and the initial service conversion feature into a multi-layer sensing network layer, perform full connection processing on the initial service triggering feature through the multi-layer sensing network layer to obtain a target service triggering feature of the target object for the target service data, and perform feature sharing transfer processing on the target service triggering feature and the initial service conversion feature in the multi-layer sensing network layer to obtain a target service conversion feature of the target object for the target service data.
It should be appreciated that the ordering sub-network further comprises a feature connection layer and a dot product full connection layer, which may be used to perform step S103 described below.
Step S103, splicing the matching features and the target service triggering features to obtain initial splicing triggering features of the target object aiming at the target service data, and determining triggering probability pre-estimation values of the target object aiming at the target service data based on the initial splicing triggering features;
specifically, the server may input the matching feature and the target service triggering feature to the feature connection layer, and perform a splicing process on the matching feature and the target service triggering feature through the feature connection layer, so as to obtain an initial splicing triggering feature of the target object for the target service data. Further, the server can input the initial splicing triggering characteristic to the dot product full-connection layer, and the dot product full-connection layer is used for carrying out full-connection processing on the initial splicing triggering characteristic to obtain the target splicing triggering characteristic. Further, the server can determine a trigger probability estimated value of the target object for the target service data according to the target splicing trigger characteristic.
The target splicing triggering characteristic is composed of a first triggering characteristic value and a second triggering characteristic value, wherein the first triggering characteristic value can represent the probability that the target service data has clicking behaviors, and the second triggering characteristic value can represent the probability that the target service data does not have clicking behaviors. Thus, the server may take the first trigger feature value as a trigger probability estimate for the target object for the target traffic data.
Step S104, the matching characteristics and the target service conversion characteristics are subjected to splicing processing to obtain initial splicing conversion characteristics of the target object aiming at the target service data, and a conversion probability estimated value of the target object aiming at the target service data is determined based on the initial splicing conversion characteristics.
Specifically, the server may input the matching feature and the target service conversion feature to the feature connection layer, and perform a splicing process on the matching feature and the target service conversion feature through the feature connection layer, so as to obtain an initial splicing conversion feature of the target object for the target service data. Further, the server can input the initial splicing transformation characteristics to the dot product full-connection layer, and the dot product full-connection layer is used for carrying out full-connection processing on the initial splicing transformation characteristics to obtain target splicing transformation characteristics. Further, the server can determine a conversion probability estimated value of the target object for the target service data according to the target splicing conversion characteristics.
The target splicing transformation characteristic is composed of a first transformation characteristic value and a second transformation characteristic value, wherein the first transformation characteristic value can represent the probability that the target service data has transformation behaviors, and the second transformation characteristic value can represent the probability that the target service data does not have transformation behaviors. Thus, the server may take the first conversion characteristic value as a conversion probability predicted value of the target object for the target service data.
Therefore, after the target service data associated with the target object is acquired, the matching characteristics between the target object and the target service data can be determined, and further the target service triggering characteristics and the target service conversion characteristics of the target object aiming at the target service data are determined. The target service conversion feature is obtained by carrying out feature sharing transfer processing on the target service triggering feature. It can be appreciated that by performing the stitching process on the matching feature and the target service triggering feature, an initial stitching triggering feature of the target object for the target service data can be obtained, where the initial stitching triggering feature can be used to generate a triggering probability pre-estimated value of the target object for the service data. Similarly, by performing stitching processing on the matching feature and the target service conversion feature, an initial stitching conversion feature of the target object for the target service data can be obtained, and the initial stitching conversion feature can be used for generating a conversion probability estimated value of the target object for the service data. Based on the above, the feature sharing transfer processing is performed on the target service triggering feature, so that the target service triggering feature corresponding to the service triggering task can be transferred to the target service conversion feature corresponding to the service conversion task, which is equivalent to sharing the bottom layer parameter corresponding to the service triggering task to the service conversion task, thereby generating the target service conversion feature with higher accuracy. In addition, the embodiment of the application can integrate the matching characteristics between the target object and the target service data into the click probability predicted value and the conversion probability predicted value, fully utilize the correlation information between the target object and the target service data, simultaneously generate the trigger probability predicted value and the conversion probability predicted value, and improve the accuracy of the trigger probability predicted value and the conversion probability predicted value of the predicted target service data.
Further, referring to fig. 4, fig. 4 is a flow chart of a data processing method according to an embodiment of the present application. The data processing method may include the following steps S1011-S1019, where steps S1011-S1019 are one embodiment of step S101 in the embodiment corresponding to fig. 3.
Step S1011, acquiring a service data set associated with the target object, and acquiring target service data from the service data set;
for a specific process of the server obtaining the service data set associated with the target object, refer to the description of step S101 in the embodiment corresponding to fig. 3, which will not be described herein.
Step S1012, inputting the object attribute of the target object, the service trigger sequence associated with the target object and the service conversion sequence associated with the target object into the feature recognition network model;
it should be understood that the service data sequence includes a service trigger sequence and a service conversion sequence. The service trigger sequence comprises service data of which the target object has trigger behaviors, and the service conversion sequence comprises service data of which the target object has conversion behaviors. The transformation behavior may include a shallow transformation behavior and a deep transformation behavior, among others. In other words, the service trigger sequence and the service conversion sequence may be collectively referred to as a service data sequence.
Alternatively, the service conversion sequence may include a shallow service conversion sequence and a deep service conversion sequence, the shallow service conversion sequence may include service data of the target object having a shallow conversion behavior, and the deep service conversion sequence may include service data of the target object having a deep conversion behavior.
It should be appreciated that the feature recognition network model includes a pre-training embedding layer, an attention processing layer, and a pre-training connection layer, which may be used to perform steps S1013-S1015 described below.
Step S1013, respectively performing feature embedding on the object attribute, the service data in the service trigger sequence and the service data in the service conversion sequence through the pre-training embedding layer to obtain an object feature corresponding to the object attribute, a trigger service feature corresponding to the service data in the service trigger sequence and a conversion service feature corresponding to the service data in the service conversion sequence;
among other things, object attributes may include, but are not limited to, object identification, base attributes, and behavior attributes. It should be understood that the server may perform a hash operation on the object identifier in the pre-training embedding layer to obtain a hashed object identifier corresponding to the object identifier. Further, the server may obtain an object identification lookup table associated with the object identification, and based on the hashed object identification, look up an object identification feature corresponding to the object identification in the object identification lookup table. Further, the server may obtain a base attribute lookup table associated with the base attribute, and find a base attribute feature corresponding to the base attribute in the base attribute lookup table. Further, the server may obtain a behavior attribute lookup table associated with the behavior attribute, and find a behavior attribute feature corresponding to the behavior attribute in the behavior attribute lookup table. Further, the server may perform feature fusion on the object identification feature, the basic attribute feature and the behavior attribute feature to obtain an object feature corresponding to the object attribute.
It may be appreciated that the object identification lookup table, the base attribute lookup table, and the behavior attribute lookup table may each include a plurality of target features (i.e., object identification features, base attribute features, behavior attribute features), where the plurality of target features in the object identification lookup table are obtained by training initial features in the initial object identification lookup table, the plurality of target features in the base attribute lookup table are obtained by training initial features in the initial base attribute lookup table, and the plurality of target features in the behavior attribute lookup table are obtained by training initial features in the initial behavior attribute lookup table. The initial object identification lookup table, the initial basic attribute lookup table and the initial behavior attribute lookup table are lookup tables in the initial probability estimation network model.
The base attributes may include, but are not limited to, the gender and age of the target subject. Therefore, the basic attribute lookup table may include an age lookup table corresponding to an age and a gender lookup table corresponding to a gender, so that the server may search for age features corresponding to the age in the age lookup table, and search for gender features corresponding to the gender in the gender lookup table, so as to perform feature fusion on the age features and the gender features, and obtain basic attribute features corresponding to the basic attributes.
Wherein the number of behavioral attributes may be one or more, where the plurality may represent at least two. It will be appreciated that when the number of behavioral attributes is one (e.g., basketball), the server may look up a behavioral attribute feature corresponding to the behavioral attribute in the behavioral attribute lookup table (i.e., look up a behavioral attribute feature corresponding to basketball). Optionally, when the number of the behavioral attributes is multiple (for example, volleyball and vegetable), the server may search the behavioral attribute features corresponding to the behavioral attributes in the behavioral attribute lookup table, and further perform feature fusion on the behavioral attribute features (that is, perform feature fusion on the behavioral attribute features corresponding to the queuing and the behavioral attribute features corresponding to the vegetable), to obtain the behavioral attribute features corresponding to the behavioral attributes.
It can be understood that the dimensions of the object identification feature, the basic attribute feature and the behavior attribute feature are the same, and the manner in which the server performs feature fusion on the object identification feature, the basic attribute feature and the behavior attribute feature may be a feature stitching manner (for example, performing feature stitching on 3 features in 128 dimensions to obtain 384 features of the object), a feature addition manner (for example, performing corresponding dimension addition on the features in 128 dimensions to obtain 128 features of the object), or a weighted average manner (for example, according to weights corresponding to the features in 128 dimensions and the 3 features, respectively, to obtain 128 features of the object). It should be understood that the embodiment of the present application does not limit the dimensions of the object identification feature, the basic attribute feature, and the behavior attribute feature, and the embodiment of the present application does not limit the specific manner in which the features are fused.
Step S1014, inputting the object feature, the triggering service feature and the converting service feature into an attention processing layer, and performing attention processing on the object feature, the triggering service feature and the converting service feature by the attention processing layer to obtain an attention object feature corresponding to the object feature, an attention triggering feature corresponding to the triggering service feature and an attention converting feature corresponding to the converting service feature;
the service data in the service trigger sequence may be referred to as trigger service data, and the service data in the service conversion sequence may be referred to as converted service data. The trigger business data corresponds to a trigger business feature and an attention trigger feature, and the conversion business data corresponds to a conversion business feature and an attention conversion feature.
The specific process of the server for performing attention processing on the object features through the attention processing layer can be seen in the following formula (1):
wherein V is user Can represent object features, V item Can represent the trigger traffic feature and the conversion traffic feature, g (V user ,V item ) Model parameters that can represent weights of users corresponding to different business data, then V' user The attention object feature may be represented and N may represent the number of triggering traffic data and translating traffic data. Wherein g (V) user ,V item ) The weight of the attention processing on the object feature may be represented, and the attention processing may be performed on the object feature based on the weight of the triggering service data and the converting service data, respectively, to obtain the attention object feature. It should be understood that, the formula of the server for performing attention processing on the trigger service feature and the conversion service feature through the attention processing layer may be referred to the above formula (1), and will not be described herein.
Step S1015, inputting the object feature, the attention triggering feature and the attention transforming feature to the pre-training connection layer, and performing feature fusion on the object feature and the attention object feature through the pre-training connection layer to obtain a pre-training object feature associated with the target object, where in the pre-training connection layer, a pre-training service feature corresponding to the service data included in the service triggering sequence and the service transforming sequence is generated according to the attention triggering feature and the attention transforming feature;
the conversion service data comprises trigger service data, the server can take service data except the trigger service data in the conversion service data as auxiliary service data, and the trigger service data and the auxiliary service data form service data contained in a service trigger sequence and a service conversion sequence. It can be understood that the server can perform feature fusion on the attention triggering feature corresponding to the triggering service data and the attention conversion feature corresponding to the triggering service data to generate the pre-training service feature corresponding to the triggering service data; the server can directly use the attention conversion characteristics corresponding to the auxiliary service data as the pre-training service characteristics corresponding to the auxiliary service data.
The specific process of performing iterative training on the initial feature recognition network model by the server to obtain the feature recognition network model can be described as: the server may obtain sample traffic data associated with the sample object, determine pre-trained sample object features associated with the sample object and pre-trained sample traffic features associated with the sample traffic data through an initial feature recognition network model. Further, the server may generate a sample quality of the sample object for the sample traffic data based on the pre-training sample object features and the pre-training sample traffic features. Further, the server may perform parameter adjustment on the initial feature recognition network model based on sample tag information and sample quality of the sample object for the sample service data, and use the initial feature recognition network model after parameter adjustment as the feature recognition network model.
The sample tag information herein may indicate whether the target object has a click action or a conversion action on the sample service data. If the target object has a click action or a conversion action for the sample service data, the sample tag information may be 1, and optionally, if the target object does not have a click action or a conversion action for the sample service data, the sample tag information may be 0.
For ease of understanding, please refer to fig. 5, fig. 5 is a schematic structural diagram of a feature recognition network model according to an embodiment of the present application. As shown in fig. 5, the server may input the object attribute of the target object, the trigger service data in the service trigger sequence (i.e. the advertisement click sequence), and the conversion service data in the service conversion sequence (i.e. the advertisement conversion sequence) to a pre-training embedding layer in the feature recognition network model, and perform feature embedding on the object attribute, the trigger service data, and the conversion service data through the pre-training embedding layer to obtain an object feature 51a corresponding to the object attribute, a trigger service feature (not shown in the figure) corresponding to the trigger service data, and a conversion service feature (not shown in the figure) corresponding to the conversion service data.
As shown in fig. 5, the server may input the object feature 51a, the trigger service feature and the conversion service feature to the attention processing layer, and perform attention processing on the object feature 51a, the trigger service feature and the conversion service feature through the attention processing layer, that is, perform attention processing on the object feature 51a, the trigger service feature and the conversion service feature based on an attention mechanism (that is, an attention mechanism), to obtain an attention object feature corresponding to the object feature 51a, an attention trigger feature corresponding to the trigger service feature and an attention conversion feature corresponding to the conversion service feature. The attention object feature, the attention triggering feature and the attention conversion feature may be the feature 51b shown in fig. 5, where the feature 51b includes multiple groups of the attention object feature, the attention triggering feature and the attention conversion feature, and the multiple groups of the attention object feature, the attention triggering feature and the attention conversion feature are obtained by performing multiple attention processes.
As shown in fig. 5, the server may input the object feature 51a and the attention object feature to a pre-training connection layer, and perform feature fusion on the object feature 51a and the attention object feature through the pre-training connection layer to obtain a pre-training object feature associated with the target object. Similarly, the server may input the attention triggering feature and the attention converting feature to the pre-training connection layer, and in the pre-training connection layer, the pre-training service feature corresponding to the service data included in the service triggering sequence and the service converting sequence is generated according to the attention triggering feature and the attention converting feature.
Step S1016, inputting the object attribute of the target object, the service attribute of the target service data, the pre-training object feature and the pre-training service feature into a matching sub-network in the probability estimation network model;
wherein the matching sub-network comprises a matching embedding layer.
Step S1017, performing feature embedding on the object attribute and the service attribute through the matching embedding layer to obtain an object feature corresponding to the object attribute and a service feature corresponding to the service attribute;
the business attributes may include, but are not limited to, business identification, business data object identification, business data category, business data text, and business data image, among others. It should be understood that the server may perform a hash operation on the service identifier in the matching embedding layer to obtain a hashed service identifier corresponding to the service identifier. Further, the server may obtain a service identifier lookup table associated with the service identifier, and based on the hashed service identifier, search the service identifier lookup table for a service identifier feature corresponding to the service identifier. Further, the server may obtain a service data object identification lookup table associated with the service data object identification, and find a service data object identification feature corresponding to the service data object identification in the service data object identification lookup table. Further, the server may obtain a service data class lookup table associated with the service data class, and find a service data class feature corresponding to the service data class in the service data class lookup table. Further, the server may obtain a service data text lookup table associated with the service data text, and look up a service data text feature corresponding to the service data text in the service data text lookup table. Further, the server may obtain a service data image lookup table associated with the service data image, and find a service data image feature corresponding to the service data image in the service data image lookup table. Further, the server may perform feature fusion on the service identifier feature, the service data object identifier feature, the service data category feature, the service data text feature, and the service data image feature, to obtain a service feature corresponding to the service attribute.
It will be appreciated that the service identification lookup table, the service data object identification lookup table, the service data category lookup table, the service data text lookup table, and the service data image lookup table may each include a plurality of target features (i.e., a service identification feature, a service data object identification feature, a service data category feature, a service data text feature, and a service data image feature), which are obtained by training the initial features. It should be understood that, for a specific process of training the initial feature to obtain the target feature, reference may be made to the description of step S1013, which will not be described herein.
It should be appreciated that, for a specific process of searching the service data class lookup table by the server, reference may be made to the above description of searching the behavior attribute lookup table, which will not be described herein. It should be appreciated that, for a specific process of searching the service data object identification lookup table by the server, reference may be made to the above description of searching the basic attribute lookup table, which will not be described herein.
The server can perform text analysis on the service data text to obtain a service data text key corresponding to the service data text, and further search service data text characteristics corresponding to the service data text in a service data text lookup table based on the service data text key. Similarly, the server can perform image analysis on the service data image to obtain a service data image key corresponding to the service data image, and further search service data image characteristics corresponding to the service data image in a service data image lookup table based on the service data image key.
It should be appreciated that, for the specific process of the server performing feature fusion on the service identification feature, the service data object identification feature, the service data category feature, the service data text feature and the service data image feature, reference may be made to the above description of performing feature fusion on the object identification feature, the basic attribute feature and the behavior attribute feature, which will not be described herein.
The specific process of the server performing feature embedding on the object attribute through the matching embedding layer can be referred to the description of performing feature embedding on the object attribute through the pre-training embedding layer, which will not be described herein.
Step S1018, in the matching sub-network, determining an object feature vector associated with the target object according to the object feature and the pre-training object feature, and determining a service feature vector associated with the target service data according to the service feature and the pre-training service feature;
for the specific process of determining the object feature vector and the service feature vector by the server, refer to the description of step S101 in the embodiment corresponding to fig. 3, which will not be described herein.
Step S1019, performing dot product operation on the object feature vector and the service feature vector to obtain the matching feature between the target object and the target service data.
Wherein the matching feature may represent a probability of matching between the target object and the target business data.
It can be seen that the embodiment of the application can generate the matching feature between the target object and the target service data according to the object attribute of the target object, the service trigger sequence associated with the target object and the service conversion sequence associated with the target object. It can be appreciated that the matching feature may be used to integrate the target business triggering feature and the target business transformation feature to integrate the correlation information between the target object and the target business data into the triggering probability predicted value and the transformation probability predicted value, thereby improving the accuracy of predicting the triggering probability predicted value and the transformation probability predicted value of the target business data.
Further, referring to fig. 6, fig. 6 is a flow chart of a data processing method according to an embodiment of the present application. The data processing method may include the following steps S1021-S1023, where steps S1021-S1023 are a specific embodiment of step S102 in the embodiment corresponding to fig. 3.
Step S1021, determining the sharing attribute characteristics of the target object aiming at the target service data through the input network layer;
specifically, the server may input the object attribute of the target object, the service attribute of the target service data, and the context attribute associated with the target object to the feature embedding layer, and perform feature embedding on the object attribute, the service attribute, and the context attribute through the feature embedding layer to obtain an object feature corresponding to the object attribute, a service feature corresponding to the service attribute, and a context feature corresponding to the context attribute. The input network layer comprises a feature embedding layer and a feature splicing layer. Further, the server may perform an average process on one or more pre-training service features to obtain an average pre-training service feature. Further, the server may input the object feature, the pre-training object feature, the service feature and the average pre-training service feature to a feature stitching layer, and in the feature stitching layer, perform feature stitching on the object feature and the pre-training object feature to obtain a stitched object feature, and perform feature stitching on the service feature and the average pre-training service feature to obtain a stitched service feature. Further, the server can perform feature stitching on the stitching object features, the stitching business features and the context features to obtain sharing attribute features of the target object aiming at the target business data.
Optionally, the server may further directly perform feature stitching on the object feature, the pre-training object feature, the service feature, the average pre-training service feature, and the context feature, to obtain a shared attribute feature of the target object for the target service data.
The service data sequence comprises one or more service data, and each service data corresponds to a pre-training service characteristic. It may be appreciated that the input network layer may further include a feature averaging layer, and the server may input one or more pre-training service features to the feature averaging layer, and average the one or more pre-training service features through the feature averaging layer to obtain an average pre-training service feature.
The context attributes may include, but are not limited to, device content and device type, among others. It should be appreciated that the server may obtain a device content look-up table associated with the device content in the feature embedding layer, and look up the device content features corresponding to the device content in the device content look-up table. Further, the server may obtain a device type lookup table associated with the device type, and look up a device type feature corresponding to the device type in the device type lookup table. Further, the server may perform feature fusion on the device content feature and the device type feature to obtain a context feature corresponding to the context attribute.
It will be appreciated that the device content look-up table and the device type look-up table may each include a plurality of target features (i.e., device content features and device type features) that are trained on the initial features. It should be understood that, for a specific process of training the initial feature to obtain the target feature, reference may be made to the description of step S1013, which will not be described herein.
It should be appreciated that, for a specific process of searching the device content lookup table by the server, reference may be made to the above description of searching the service data text lookup table, which will not be described herein. It should be appreciated that, for a specific process of searching the device content type lookup table by the server, reference may be made to the above description of searching the basic attribute lookup table, which will not be described herein.
The specific process of the server performing feature embedding on the object attribute through the feature embedding layer can be referred to the description of performing feature embedding on the object attribute through the pre-training embedding layer, and will not be described in detail herein; the specific process of the server performing feature embedding on the service attribute through the feature embedding layer can be referred to the description of performing feature embedding on the service attribute through the pre-training embedding layer, which will not be described herein.
The dimensions of the object feature, the service feature, the context feature, the pre-training object feature and the average pre-training service feature may be N dimensions, where N may be a positive integer, and the dimension of the shared attribute feature is 5N (i.e., 5*N) dimensions. It should be understood that the embodiment of the present application does not limit the value of N.
It should be appreciated that the server may initialize the initial features in the lookup table (i.e., the unbedding table) prior to training the initial probability predictive network model, and may further update the initial features in the lookup table according to the inverse gradient during training of the initial probability predictive network model to obtain the lookup table containing the target features. In addition, for high-dimensional features (e.g., service identification features), the server may map the high-latitude features to low-dimensional features before storing the low-dimensional features in a lookup table.
Step S1022, inputting the shared attribute feature into the parameter sharing network layer, and generating an initial service triggering feature and an initial service conversion feature of the target object aiming at the target service data through the parameter sharing network layer;
specifically, the server may input the shared attribute feature to the shared full-connection layer, and perform full-connection processing on the shared attribute feature through the shared full-connection layer to obtain a full-connection shared feature. The parameter sharing network layer comprises a sharing full-connection layer, a weight learning layer, a feature classification layer, a first triggering full-connection layer and an initial conversion full-connection layer. Further, the server may input the full-connection sharing feature to the weight learning layer, and perform feature weighting processing on the full-connection sharing feature through the weight learning layer to obtain a weight sharing feature. Further, the server may input the weight sharing feature to a feature classification layer, and perform feature classification on the weight sharing feature through the feature classification layer to obtain a trigger distribution vector and a conversion distribution vector of the target object for the target service data. Further, the server may input the trigger distribution vector to the first trigger full-connection layer, and perform full-connection processing on the trigger distribution vector through the first trigger full-connection layer, so as to generate an initial service trigger feature of the target object for the target service data. Further, the server may input the transformation distribution vector to an initial transformation full connection layer, and perform full connection processing on the transformation distribution vector through the initial transformation full connection layer, so as to generate an initial service transformation characteristic of the target object for the target service data.
The weight learning layer comprises weight learning components (namely Gate) corresponding to object features, service features, context features, pre-training object features and average pre-training service features respectively; the feature classification layer comprises a trigger feature classification layer and a conversion feature classification layer. Therefore, the server can input the full-connection sharing feature into each weight learning component respectively, and the feature weighting processing is carried out on the full-connection sharing feature through each weight learning component respectively to obtain the weight sharing feature corresponding to each weight learning component. Further, the server may perform feature fusion on the weight sharing feature corresponding to each weight learning component, to obtain a fused sharing feature. Further, the server may input the fusion shared feature to a trigger feature classification layer, and perform feature classification on the fusion shared feature through the trigger feature classification layer to obtain a trigger distribution vector of the target object for the target service data. Further, the server can input the fusion shared features to a transformation feature classification layer, and the transformation feature classification layer performs feature classification on the fusion shared features to obtain transformation distribution vectors of the target object aiming at the target service data.
Step S1023, inputting the initial service triggering characteristic and the initial service conversion characteristic into a multi-layer sensing network layer, performing full connection processing on the initial service triggering characteristic through the multi-layer sensing network layer to obtain the target service triggering characteristic of the target object aiming at the target service data, and performing characteristic sharing transmission processing on the target service triggering characteristic and the initial service conversion characteristic in the multi-layer sensing network layer to obtain the target service conversion characteristic of the target object aiming at the target service data.
Specifically, the server may input the initial service triggering feature to the second triggering full-connection layer, and perform full-connection processing on the initial service triggering feature through the second triggering full-connection layer, so as to obtain a target service triggering feature of the target object for the target service data. The multi-layer perception network layer comprises a second triggering full-connection layer, a target conversion full-connection layer and a conversion connection layer. Further, the server may input the target service triggering feature and the initial service conversion feature to the conversion connection layer, and perform feature stitching on the target service triggering feature and the initial service conversion feature through the conversion connection layer to obtain a shared stitching feature. Further, the server can input the shared splicing feature to a target conversion full-connection layer, and full-connection processing is performed on the shared splicing feature through the target conversion full-connection layer, so that target business conversion features of the target object aiming at target business data are obtained.
Therefore, the embodiment of the application can determine the sharing attribute characteristics of the target object aiming at the target service data through the sequencing subnetwork, and further generate the target service triggering characteristics and the target service conversion characteristics of the target object aiming at the target service data based on the sharing attribute characteristics. It can be understood that the target service conversion feature is obtained by performing shared transfer processing on the target service triggering feature, and when the triggering probability predicted value is generated through the target service triggering feature and the conversion probability predicted value is generated through the target service conversion feature, the accuracy of the triggering probability predicted value and the conversion probability predicted value of the predicted target service data can be improved.
Further, referring to fig. 7, fig. 7 is a flow chart of a data processing method according to an embodiment of the present application. The method may be executed by a server, or may be executed by an application client, or may be executed by a server and an application client together, where the server may be the server 20a in the embodiment corresponding to fig. 2, and the application client may be the application client in the embodiment corresponding to fig. 2. For ease of understanding, embodiments of the present application will be described in terms of this method being performed by a server. The data processing method may include the following steps S201 to S207:
Step S201, acquiring target business data associated with a target object, and determining matching characteristics between the target object and the target business data;
for a specific process of determining the matching feature between the target object and the target service data by the server, refer to the description of step S101 in the embodiment corresponding to fig. 3, which will not be described herein.
Step S202, determining the sharing attribute characteristics of a target object aiming at target service data through an input network layer;
for a specific process of determining the shared attribute feature by the server, refer to the description of step S1021 in the embodiment corresponding to fig. 6, which will not be described in detail herein.
Step S203, the shared attribute feature is input to a parameter sharing network layer, and an initial service triggering feature and an initial service conversion feature of a target object aiming at target service data are generated through the parameter sharing network layer;
specifically, the server may input the shared attribute feature to the shared full-connection layer, and perform full-connection processing on the shared attribute feature through the shared full-connection layer to obtain a full-connection shared feature. The parameter sharing network layer comprises a sharing full-connection layer, a weight learning layer, a feature classification layer, a first triggering full-connection layer, a first shallow conversion full-connection layer and a first deep conversion full-connection layer; the initial business transformation features include an initial shallow transformation feature and an initial deep transformation feature. Further, the server may input the full-connection sharing feature to the weight learning layer, and perform feature weighting processing on the full-connection sharing feature through the weight learning layer to obtain a weight sharing feature. Further, the server may input the weight sharing feature to a feature classification layer, and perform feature classification on the weight sharing feature through the feature classification layer to obtain a trigger distribution vector, a shallow conversion distribution vector and a deep conversion distribution vector of the target object for the target service data. Further, the server may input the trigger distribution vector to the first trigger full-connection layer, and perform full-connection processing on the trigger distribution vector through the first trigger full-connection layer, so as to generate an initial service trigger feature of the target object for the target service data. Further, the server may input the shallow conversion distribution vector to the first shallow conversion full connection layer, and perform full connection processing on the shallow conversion distribution vector through the first shallow conversion full connection layer, so as to generate an initial shallow conversion feature of the target object for the target service data. Further, the server may input the deep conversion distribution vector to the first deep conversion full-connection layer, and perform full-connection processing on the deep conversion distribution vector through the first deep conversion full-connection layer, so as to generate an initial deep conversion feature of the target object for the target service data.
The weight learning layer comprises weight learning components corresponding to object features, business features, context features, pre-training object features and average pre-training business features respectively; the feature classification layer comprises a trigger feature classification layer, a shallow conversion feature classification layer and a deep conversion feature classification layer. Therefore, the server can input the full-connection sharing feature into each weight learning component respectively, and the feature weighting processing is carried out on the full-connection sharing feature through each weight learning component respectively to obtain the weight sharing feature corresponding to each weight learning component. Further, the server may perform feature fusion on the weight sharing feature corresponding to each weight learning component, to obtain a fused sharing feature. Further, the server may input the fusion shared feature to a trigger feature classification layer, and perform feature classification on the fusion shared feature through the trigger feature classification layer to obtain a trigger distribution vector of the target object for the target service data. Further, the server can input the fusion shared feature to a shallow conversion feature classification layer, and the shallow conversion feature classification layer is used for classifying the fusion shared feature to obtain a shallow conversion distribution vector of the target object aiming at the target service data. Further, the server may input the fusion shared feature to a deep conversion feature classification layer, and perform feature classification on the fusion shared feature through the deep conversion feature classification layer to obtain a deep conversion distribution vector of the target object for the target service data.
Step S204, inputting the initial service triggering characteristic and the initial service conversion characteristic into a multi-layer sensing network layer, performing full connection processing on the initial service triggering characteristic through the multi-layer sensing network layer to obtain a target service triggering characteristic of a target object aiming at target service data, and performing characteristic sharing transmission processing on the target service triggering characteristic and the initial service conversion characteristic in the multi-layer sensing network layer to obtain a target service conversion characteristic of the target object aiming at the target service data;
specifically, the server may input the initial service triggering feature to the second triggering full-connection layer, and perform full-connection processing on the initial service triggering feature through the second triggering full-connection layer, so as to obtain a target service triggering feature of the target object for the target service data. The multi-layer sensing network layer comprises a second triggering full-connection layer, a second shallow conversion full-connection layer, a shallow connection layer, a second deep conversion full-connection layer and a deep connection layer; the initial business transformation features comprise initial shallow transformation features and initial deep transformation features; the target business transformation features include target shallow transformation features and target deep transformation features. Further, the server can input the target service triggering characteristic and the initial shallow conversion characteristic to the shallow connecting layer, and the target service triggering characteristic and the initial shallow conversion characteristic are subjected to characteristic splicing through the shallow connecting layer to obtain the shared shallow splicing characteristic. Further, the server may input the shared shallow splicing feature to a second shallow conversion full-connection layer, and perform full-connection processing on the shared shallow splicing feature through the second shallow conversion full-connection layer, so as to obtain a target shallow conversion feature of the target object for the target service data. Further, the server may input the target shallow layer transformation feature and the initial deep layer transformation feature to the deep layer connection layer, and perform feature stitching on the target shallow layer transformation feature and the initial deep layer transformation feature through the deep layer connection layer to obtain a shared deep layer stitching feature. Further, the server may input the shared deep splicing feature to a second deep conversion full-connection layer, and perform full-connection processing on the shared deep splicing feature through the second deep conversion full-connection layer, so as to obtain a target deep conversion feature of the target object for the target service data.
Step 205, performing stitching processing on the matching feature and the target service triggering feature to obtain an initial stitching triggering feature of the target object for the target service data, and determining a triggering probability estimated value of the target object for the target service data based on the initial stitching triggering feature;
for a specific process of determining the trigger probability estimated value by the server, refer to the description of step S103 in the embodiment corresponding to fig. 3, which will not be repeated here.
Wherein the sequencing subnetwork further comprises a feature connection layer and a dot product full connection layer, which can be used to perform step S206 and step S207 described below.
Step 206, performing splicing processing on the matching features and the target shallow conversion features to obtain initial spliced shallow conversion features of the target object aiming at the target service data, and determining shallow conversion probability pre-estimation values of the target object aiming at the target service data based on the initial spliced shallow conversion features;
specifically, the server may input the matching feature and the target shallow conversion feature to the feature connection layer, and perform a splicing process on the matching feature and the target shallow conversion feature through the feature connection layer, so as to obtain an initial spliced shallow conversion feature of the target object for the target service data. Further, the server can input the initial splicing shallow layer conversion characteristics to the dot product full-connection layer, and full-connection processing is carried out on the initial splicing shallow layer conversion characteristics through the dot product full-connection layer to obtain target splicing shallow layer conversion characteristics. Further, the server can determine a shallow conversion probability estimated value of the target object for the target service data according to the target splicing shallow conversion characteristics.
The target splicing shallow conversion characteristic is composed of a first shallow conversion characteristic value and a second shallow conversion characteristic value, wherein the first shallow conversion characteristic value can represent the probability that target service data has shallow conversion behaviors, and the second shallow conversion characteristic value can represent the probability that the target service data does not have shallow conversion behaviors. Therefore, the server can take the first shallow conversion characteristic value as a shallow conversion probability pre-estimated value of the target object aiming at the target service data.
Step 207, performing stitching processing on the matching feature and the target deep layer transformation feature to obtain an initial stitching deep layer transformation feature of the target object for the target service data, and determining a deep layer transformation probability estimated value of the target object for the target service data based on the initial stitching deep layer transformation feature.
Specifically, the server may input the matching feature and the target deep conversion feature to the feature connection layer, and perform a stitching process on the matching feature and the target deep conversion feature through the feature connection layer, so as to obtain an initial stitching deep conversion feature of the target object for the target service data. Further, the server can input the initial splicing deep conversion characteristics to the dot product full-connection layer, and the full-connection processing is carried out on the initial splicing deep conversion characteristics through the dot product full-connection layer to obtain target splicing deep conversion characteristics. Further, the server can determine a deep conversion probability estimated value of the target object for the target service data according to the target splicing deep conversion characteristics.
The target splicing deep layer transformation characteristic is composed of a first deep layer transformation characteristic value and a second deep layer transformation characteristic value, wherein the first deep layer transformation characteristic value can represent the probability that target service data has deep layer transformation behaviors, and the second deep layer transformation characteristic value can represent the probability that the target service data does not have deep layer transformation behaviors. Therefore, the server can take the first deep conversion characteristic value as a deep conversion probability estimated value of the target object aiming at the target service data.
It should be appreciated that the transition probability predictors include shallow transition probability predictors and deep transition probability predictors; the target business transformation features include target shallow transformation features and target deep transformation features. In other words, the shallow conversion probability predictive value and the deep conversion probability predictive value may be collectively referred to as conversion probability predictive values, and the target shallow conversion feature and the target deep conversion feature may be collectively referred to as target business conversion feature.
It can be understood that, the specific process of determining the trigger probability estimated value, the shallow conversion probability estimated value and the deep conversion probability estimated value of the target object for the service data by the server may be referred to fig. 8a and 8b, where fig. 8a and 8b are schematic structural diagrams of a probability estimation network model provided in the embodiments of the present application, fig. 8a corresponds to a model structure of a matching sub-network, and fig. 8b corresponds to a model structure of a sorting sub-network.
As shown in fig. 8a, the matching sub-network may be in a dual-tower structure, and the server may perform feature embedding on the object attribute and the service attribute through the matching embedding layer to obtain an object feature corresponding to the object attribute and a service feature corresponding to the service attribute, further input the object feature to the trigger full-connection layer, perform full-connection processing on the object feature through the trigger full-connection layer to obtain a full-connection object feature, input the service feature to the service full-connection layer, and perform full-connection processing on the service feature through the service full-connection layer to obtain a full-connection service feature.
As shown in fig. 8a, the server may input the pre-training service features to a feature averaging layer, and average the pre-training service features through the feature averaging layer to obtain average pre-training service features. Further, the server may perform feature fusion on the pre-training object feature and the full-connection object feature to obtain an object feature vector, and perform feature fusion on the average pre-training service feature and the full-connection service feature to obtain a service feature vector. Further, the server may perform a dot product operation on the service feature vector and the object feature vector to obtain a matching feature between the target object and the target service data.
As shown in fig. 8b, the server may input the object attribute, the service attribute and the context attribute to the input network layer 80b, and perform feature embedding on the object attribute, the service attribute and the context attribute through the feature embedding layer in the input network layer 80b to obtain an object feature corresponding to the object attribute, a service feature corresponding to the service attribute and a context feature corresponding to the context attribute. Further, the server may input the pre-training service feature to a feature average layer in the input network layer 80b, output the pre-training service feature through the feature average layer, further perform feature stitching on the pre-training service feature and the service feature to obtain a stitched service feature, and perform feature stitching on the object feature and the pre-training object feature to obtain a stitched object feature. Further, the server may perform feature stitching on the stitched object features, the stitched service features, and the context features through a feature stitching layer (not shown in the figure, for convenience of understanding, and shown in the figure by taking the feature embedding layer and the feature stitching layer as the same layer) in the input network layer 80b, so as to obtain the shared attribute feature of the target object for the target service data.
As shown in fig. 8b, the server may input the shared attribute feature output by the input network layer 80b to the parameter sharing network layer 80c, and output the full connection shared feature by the shared full connection layer in the parameter sharing network layer 80 c. Further, the server may input the fully connected shared features to the weight learning component 1, the weight learning component 2, the weight learning component 3, the weight learning component 4 and the weight learning component 5 in the shared network layer 80c, respectively, and output the weight shared features corresponding to the weight learning components through the weight learning components, so as to perform feature fusion on the weight shared features corresponding to the weight learning component 1, the weight learning component 2, the weight learning component 3, the weight learning component 4 and the weight learning component 5, respectively, to obtain fused shared features, so that the fused shared features are input to the trigger feature classification layer, the shallow conversion feature classification layer and the deep conversion feature classification layer in the parameter shared network layer 80c, respectively. Wherein the weight learning component 1 may correspond to object features, the weight learning component 2 may correspond to business features, the weight learning component 3 may correspond to contextual features, the weight learning component 4 may correspond to pre-trained object features, and the weight learning component 5 may correspond to pre-trained business features. Further, the server may input the trigger distribution vector corresponding to the trigger feature classification layer to the first trigger full-connection layer in the parameter sharing network layer 80c, input the shallow conversion distribution vector corresponding to the shallow conversion feature classification layer to the first shallow conversion full-connection layer in the parameter sharing network layer 80c, input the deep conversion corresponding to the deep conversion feature classification layer to the first deep conversion full-connection layer in the parameter sharing network layer 80c, output the initial service trigger feature by the first trigger full-connection layer, output the initial shallow conversion feature by the first shallow conversion full-connection layer, and output the initial deep conversion feature by the first deep conversion full-connection layer. Wherein the first shallow conversion full-link layer and the first deep conversion full-link layer shown in fig. 8b may be collectively referred to as an initial conversion full-link layer.
As shown in fig. 8b, the server may input the initial traffic trigger feature to the multi-layer aware network layer 80d and output the target traffic trigger feature by a second trigger fully connected layer in the multi-layer aware network layer 80 d. Further, the server may perform feature stitching on the target service triggering feature and the initial shallow conversion feature through the shallow connection layer in the multi-layer sensing network layer 80d to obtain a shared shallow stitching feature, and then output the target shallow conversion feature through the second shallow conversion full connection layer in the multi-layer sensing network layer 80 d. Further, the server may perform feature stitching on the target shallow layer transformation feature and the initial deep layer transformation feature through the deep layer connection layer in the multi-layer perception network layer 80d to obtain a shared deep layer stitching feature, and then output the target deep layer transformation feature through the second deep layer transformation full connection layer in the multi-layer perception network layer 80 d.
As shown in fig. 8b, the server may acquire a matching feature between the target object and the target service data, perform a splicing process on the matching feature and the target service triggering feature through the feature connection layer to obtain an initial splicing triggering feature, and further output a triggering probability pre-estimated value of the target object for the target service data through the dot product full connection layer; splicing the matching features and the target shallow conversion features through the feature connection layer to obtain initial spliced shallow conversion features, and outputting shallow conversion probability estimated values of the target object aiming at the target service data through the dot product full connection layer; and performing splicing treatment on the matching features and the target deep conversion features through the feature connection layer to obtain initial spliced deep conversion features, and outputting a deep conversion probability predicted value of the target object aiming at the target service data through the dot product full connection layer. Wherein the dot product full connectivity layer may use a sigmod function as the activation function.
It should be understood that the server may perform operation processing on the trigger probability predicted value, the conversion probability predicted value, and the predicted resource corresponding to the target service data, to obtain the service data quality of the target object for the target service data. Wherein the estimated resources may represent shallow target bids and/or deep target bids for targeted traffic data by advertisers, traffic data quality may represent ECPM (Effective Cost Per Mille) metrics, and ECPM represents advertising revenue available per thousand impressions of an advertisement. Further, the server may perform a first sorting process on the target service data based on the quality of the service data, to obtain initial sorting information corresponding to the target service data. Further, the server can perform second sorting processing on the target service data through the sorting strategy and the initial sorting information to obtain target sorting information corresponding to the target service data.
The method of the server performing operation processing on the trigger probability estimated value, the conversion probability estimated value and the estimated resource can be multiplication operation, that is, the server can take the product of the trigger probability estimated value, the conversion probability estimated value and the estimated resource as the quality of service data.
It will be appreciated that the server, upon receiving a service data acquisition request sent by the target object, may determine a matching probability between the target object and the target service data (i.e. a matching score log of the target object for the target service data 0 ) Trigger probability predictive value (i.e. click rate predictive value ectr (Expect Click Through Rate)) of target object aiming at target service data, shallow conversion probability predictive value (i.e. shallow conversion rate predictive value ecvr) 1 (Expect Conversion Rate 1)) and a deep-layer conversion probability predictive value (i.e., a deep-layer conversion predictive value ecvr) 2 (Expect Conversion Rate 2)). Wherein, the click rate predicted value (click rate for short) can be represented by the following formula (2), the shallow conversion rate predicted value (shallow conversion rate for short) can be represented by the following formula (3), and the deep conversion rate predicted value (deep conversion rate for short) can be represented by the following formula (4):
ectr=P(ectr|x,H,logit 0 ) (2)
ecvr 1 =P(ecvr 1 |ectr,x,H,logit 0 ) (3)
ecvr 2 =P(ecvr 2 |ectr,ecvr 1 ,x,H,logit 0 ) (4)
where x may represent all features of the input (i.e., object attributes, business attributes, context attributes, trigger business features, and conversion business features), and H may represent parameters of the probabilistic predictive network model. Thus, P (ctr|x, H, logic) 0 ) Can represent the click rate predicted value, P (ecvr) 1 |ectr,x,H,logit 0 ) Shallow conversion predicted value, P (ecvr) 2 |ectr,ecvr 1 ,x,H,logit 0 ) Can represent a probabilistic predictive network modelAnd outputting a predicted value of the deep conversion rate.
It will be appreciated that where traffic data pertains to shallow targeted advertising, the manner in which the server determines the quality of the traffic data may be found in equation (5) below:
ecpm=ectr*ecvr 1 *bid 1 (5)
wherein, ctr can represent the estimated click probability (i.e. click rate pre-estimated value) of the single-target advertisement, ecvr 1 May represent the estimated shallow conversion probability (i.e., shallow conversion predictive value) of a single targeted advertisement, bid 1 Shallow target bids for single target advertisements may be represented.
Optionally, when the service data belongs to the deep targeted advertisement, if the shallow conversion rate predicted value is greater than the deep conversion rate predicted value, the server may determine the quality of the service data according to the following formula (6), and if the shallow conversion rate predicted value is not greater than (i.e. less than or equal to) the deep conversion rate predicted value, the server may determine the quality of the service data according to the following formula (7):
ecpm=ectr*ecvr 1 *bid 1 (6)
ecpm=ectr*ecvr 2 *bid 2 (7)
wherein, ctr can represent the estimated click probability (i.e. click rate pre-estimated value) of the multi-target advertisement, ecvr 1 The estimated shallow conversion probability (i.e., shallow conversion rate predicted value) of the multi-target advertisement can be expressed, ecvr 2 The estimated deep conversion probability (i.e., the predicted deep conversion rate) of the multi-target advertisement can be represented, bid 1 Shallow target bids, bid, which may represent multi-target advertising 2 Deep targeted bids for multi-targeted advertising may be represented.
For ease of understanding, please refer to fig. 9, fig. 9 is a schematic flow chart of service data recommendation provided in the embodiment of the present application. As shown in fig. 9, the application client may initiate a request (i.e., a service data acquisition request) to the server through step S91, so that the server may retrieve a recall advertisement from the advertisement library through step S92, where the advertisement library may be the service database 21a in the embodiment corresponding to fig. 2, and the recall advertisement may be service data (i.e., an advertisement) in the service data set acquired through the recall model.
As shown in fig. 9, the server may perform step S93, in which the click rate of the advertisement is estimated through the probability estimation network model, and further perform step S94, and the service type of the service data is determined through step S94, where the service type may include a single-target advertisement and a multi-target advertisement. It can be understood that, when the service data is a single-target advertisement, the server may execute step S95, in step S95, the shallow conversion rate of the advertisement is estimated through the probability estimation network model, and then execute step S96, to calculate the quality (i.e. ECPM) of the single-target advertisement based on the click rate and the shallow conversion rate of the single-target advertisement. Optionally, it may be understood that when the service data is a multi-target advertisement, the server may execute step S97, in step S97, the shallow conversion rate and the deep conversion rate of the advertisement are estimated through the probability estimation network model, and then execute step S98, to calculate the quality of the multi-target advertisement based on the click rate, the shallow conversion rate and the deep conversion rate of the multi-target advertisement.
Therefore, after the quality of the single-target advertisement is calculated by the server, the single-target advertisement can be sequenced based on the quality of the single-target advertisement, so that initial sequencing information of the single-target advertisement is obtained; after the server calculates the quality of the multi-target advertisement, the server may sort the multi-target advertisement based on the quality of the multi-target advertisement to obtain initial sorting information of the multi-target advertisement.
Further, as shown in fig. 9, in step S99, the server may reorder the advertisements according to ordering policies such as advertisement diversity, frequency control, category control, specific result weighting, etc., to obtain target ordering information of the single-target advertisement or the multi-target advertisement, further execute step S100, obtain the advertisement with the first final rank from the target ordering information through step S100, and expose the advertisement (i.e. recommended service data) obtained from the target ordering information to the user (i.e. target object) on the medium (i.e. multimedia data displayed in the application client). Wherein, advertisement diversity, frequency control, category control may be collectively referred to as ranking policy. Optionally, the server may also obtain the advertisement with the first final rank from the initial ranking information without considering the ranking policy, and expose the advertisement (i.e., the recommended service data) obtained from the initial ranking information to the target object on the medium.
It can be seen that embodiments of the present application may include three tasks: one is a click rate estimation task, one is a shallow conversion rate estimation task, and the other is a deep conversion rate estimation task. The sharing mechanism can enable the shallow conversion rate estimation task and the deep conversion rate estimation task to share the bottom parameters of the click rate estimation task, namely, the target service triggering characteristics of the click rate estimation task are shared to the conversion rate estimation task (namely, the shallow conversion rate estimation model and the deep conversion rate estimation model), specifically, the target service triggering characteristics of the click rate estimation task are shared to the shallow conversion rate estimation task, and the target shallow conversion characteristics of the shallow conversion rate estimation task are shared to the deep conversion rate estimation task, so that the accuracy of the estimated triggering probability pre-estimation value, the shallow conversion probability pre-estimation value and the deep conversion probability pre-estimation value can be improved. Therefore, when the triggering probability predicted value and the conversion probability predicted value with higher accuracy are used for sequencing the target service data, the sequencing information with more accurate target service data can be obtained, and further when the sequencing information is used for recommending the service data for the target object, the accuracy of service data recommendation can be improved, similar advertisements are prevented from being repeatedly recommended, and the user experience of an advertisement recommendation system (namely the service data recommendation system) is improved.
Further, referring to fig. 10, fig. 10 is a flow chart of a data processing method according to an embodiment of the present application. The method may be executed by a server, or may be executed by an application client, or may be executed by a server and an application client together, where the server may be the server 20a in the embodiment corresponding to fig. 2, and the application client may be the application client in the embodiment corresponding to fig. 2. For ease of understanding, embodiments of the present application will be described in terms of this method being performed by a server. The data processing method may include the following steps S301 to S305:
step S301, sample service data associated with a sample object is obtained, and sample matching characteristics between the sample object and the sample service data are determined through an initial probability pre-estimated network model;
wherein the sample object has a click action, a shallow conversion action or a deep conversion action for sample business data, and the number of the sample business data is one or more.
It should be appreciated that, for a specific process of determining the matching feature of the sample by the server through the initial probability estimation network model, reference may be made to the description of determining the matching feature by the probability estimation network model, which will not be described herein.
Step S302, in an initial probability pre-estimated network model, determining sample service triggering characteristics and sample service conversion characteristics of a sample object aiming at sample service data;
the sample service conversion characteristics are obtained by carrying out characteristic sharing transfer processing on the sample service triggering characteristics, and the sample service conversion characteristics comprise sample shallow conversion characteristics and sample deep conversion characteristics.
The specific process of determining the sample service triggering feature and the sample service conversion feature by the server through the initial probability prediction network model can be referred to the description of determining the target service triggering feature and the target service conversion feature through the probability prediction network model, which will not be described in detail herein.
Step S303, performing splicing processing on the sample matching features and the sample service triggering features to obtain sample splicing triggering features of the sample objects aiming at the sample service data, and determining sample triggering probability pre-estimation values of the sample objects aiming at the sample service data based on the sample splicing triggering features;
it should be appreciated that, for a specific process of determining the trigger probability estimated value of the sample by the server through the initial probability estimation network model, reference may be made to the description of determining the trigger probability estimated value by the probability estimation network model, which will not be described herein.
Step S304, performing splicing processing on the sample matching characteristics and the sample service conversion characteristics to obtain sample splicing conversion characteristics of the sample object aiming at the sample service data, and determining a sample conversion probability estimated value of the sample object aiming at the sample service data based on the sample splicing conversion characteristics;
the sample transformation probability pre-estimation value comprises a sample shallow transformation probability pre-estimation value and a sample deep transformation probability pre-estimation value, and the sample splicing transformation characteristic comprises a sample shallow splicing transformation characteristic and a sample deep splicing transformation characteristic.
It should be appreciated that, for a specific process of determining the conversion probability estimated value of the sample by the server through the initial probability estimation network model, reference may be made to the description of determining the conversion probability estimated value by the probability estimation network model, which will not be described herein.
Step S305, carrying out parameter adjustment on the initial probability prediction network model based on sample label information, sample triggering probability prediction value and sample conversion probability prediction value of the sample object aiming at sample service data, and taking the initial probability prediction network model after parameter adjustment as a probability prediction network model;
specifically, the server may determine a trigger loss value of the initial probability pre-estimated network model based on the service trigger tag information and the sample trigger probability pre-estimated value of the sample object for the sample service data. The sample tag information comprises service trigger tag information and service conversion tag information. Further, the server may determine a conversion loss value of the initial probability prediction network model based on the sample object for the sample service data service conversion label information and the sample conversion probability prediction value. Further, the server may determine a model loss value for the initial probabilistic predictive network model based on the trigger loss value and the conversion loss value. Further, the server can perform parameter adjustment on the initial probability prediction network model according to the model loss value, and when the initial probability prediction network model after parameter adjustment meets the model convergence condition, the initial probability prediction network model after parameter adjustment is used as the probability prediction network model. The probability prediction network model is used for predicting a trigger probability predicted value and a conversion probability predicted value of the target object aiming at the target service data.
The business transformation tag information comprises shallow transformation tag information and deep transformation tag information. It can be appreciated that the server can determine the shallow conversion loss value of the initial probability predictive network model based on the shallow conversion label information of the sample object for the sample traffic data and the sample shallow conversion probability predictive value. Further, the server may determine a deep conversion loss value of the initial probability pre-estimated network model based on the deep conversion label information of the sample object for the sample service data and the sample deep conversion probability pre-estimated value. Further, the server may determine a conversion loss value of the initial probability predictive network model based on the shallow conversion loss value and the deep conversion loss value.
It should be appreciated that the specific process of the server performing parameter adjustment (i.e., iterative training) on the initial probability estimation network model according to the model loss value can be described as: when the model loss value of the initial probability pre-estimated network model does not meet the model convergence condition, the server can adjust the model parameters of the initial probability pre-estimated network model based on the model loss value which does not meet the model convergence condition. Further, the server may determine the initial probability estimation network model after the model parameter adjustment as a transition probability estimation network model, and perform iterative training on the transition probability estimation network model, until the model loss value of the transition probability estimation network model after the iterative training meets the model convergence condition, and take the transition probability estimation network model meeting the model convergence condition as the probability estimation network model.
It should be appreciated that the initial probability prediction network model and the probability prediction network model may be collectively referred to as a generalized probability prediction network model, where the initial probability prediction network model and the probability prediction network model belong to names of the generalized probability prediction network model at different moments, and in the training stage, the generalized probability prediction network model may be referred to as an initial probability prediction network model, and in the prediction stage, the generalized probability prediction network model may be referred to as a probability prediction network model.
It can be appreciated that when the server performs iterative training on the initial probability estimation network model, the server can output probability values P (ctr, ecvr) of a plurality of targets through the initial probability estimation network model 1 ,ecvr 2 The server can decompose the joint distribution into smaller single distribution, and obtain the probability value output by the initial probability prediction network model according to the assumption of the Bayesian probability model. Wherein, the probability value output by the initial probability estimating network model can be referred to the following formula (8):
where x may represent all features of the input (i.e., sample object attributes, sample traffic attributes, and sample context attributes), and H may represent parameters of the initial probabilistic predictive network model. Thus, P (ctr|x, H, logic) 0 ) Can represent the estimated value of the sample conversion probability, P (ecvr) 1 |ectr,x,H,logit 0 ) Can represent the sample shallow conversion probability predictive value, P (ecvr) 2 |ectr,ecvr 1 ,x,H,logit 0 ) The method can represent the sample deep conversion probability estimated value output by the target probability estimated network model.
It will be appreciated that taking a negative log-likelihood loss function (i.e., negative log-likelihood) for equation (8) results in a loss function L (x, H) shown in equation (9) below:
it will be appreciated that adding weight parameters (weights for short) to different targets in equation (9) can result in a loss function L (x, H) containing weight parameters as shown in equation (10) below:
wherein W is 1 、W 2 And W is 3 Weights of three loss functions, W, can be represented respectively 1 Representing the loss function P (ctr|x, H, logic 0 ) Corresponding weight, W 2 Can be represented as P (ecvr 1 |ectr,x,H,logit 0 ) Corresponding weight, W 3 Can be represented as P (ecvr 2 |ectr,ecvr 1 ,x,H,logit 0 ) And (5) corresponding weight.
It is understood that the weight parameter may represent a click behavior, a shallow transformation behavior, or a deep transformation behavior of the sample object with respect to the sample business data. For example, when the sample object has click behavior, no shallow conversion behavior and no deep conversion behavior with respect to the sample business data, W 1 Can be equal to 1, W 2 And W is 3 May be equal to 0; for another example, when the sample object has click behavior and shallow conversion behavior for sample business data, and does not have deep conversion behavior, W 1 And W is 2 Can be equal to 1, W 3 May be equal to 0; for another example, when the sample object has a click behavior, a shallow conversion behavior, and a deep conversion behavior with respect to the sample business data, W 1 、W 2 And W is 3 May be equal to 1.
It should be appreciated that the initial probability estimation network model may be composed of three parts as shown in FIG. 8b, one part may be used to fit ctr and another part may be used to fit ecvr as shown in FIG. 8b 1 Yet another part can be used to fit ecvr 2 . Therefore, the initial probability estimation network model has three subtasks, and the three subtasks can be respectively used for outputting ctr and ecvr 1 And ecvr 2 Because embodiments of the present application may predict multiple objectives based on a bayesian network: electric, ecvr 1 And ecvr 2 The goal of the server training the initial probability estimation network model is to optimize P (ctr, ecvr 1 ,ecvr 2 |x,H,logit 0 )。
Therefore, the embodiment of the application provides a multitasking model based on a pre-trained Bayesian network, which integrates the matching score, click rate, shallow conversion rate and deep conversion rate of a target object and target business data into one model, integrates the intermediate training result of the click rate into the training of the shallow conversion rate target, integrates the intermediate training result of the shallow conversion rate into the training of the deep conversion rate target, integrates the correlation information between the target object and the target business data into the training of the click rate, the shallow conversion rate and the deep conversion rate, and integrates the matching score into the training of the click rate, the shallow conversion rate and the deep conversion rate. Based on the method, three targets of click rate, shallow conversion rate and deep conversion rate are used at the same time, so that the initial probability estimation network model can fully utilize bottom parameters, the problem of sparse training data is relieved, the training samples of click rate estimation of the click rate estimation model are shared to the conversion rate estimation model, the shallow conversion samples of the shallow conversion rate estimation model are shared to the deep conversion rate estimation model (in practice, the click rate estimation model, the conversion rate estimation model, the shallow conversion rate estimation model and the deep conversion rate estimation model all belong to the initial probability estimation network model), and therefore the accuracy of click rate, shallow conversion rate and deep conversion rate estimation can be improved, and the accuracy of service data recommendation is improved.
Further, referring to fig. 11, fig. 11 is a schematic structural diagram of a data processing apparatus according to an embodiment of the present application. The data processing apparatus 1 may include: a first determining module 11, a second determining module 12, a first estimating module 13, a second estimating module 14;
a first determining module 11, configured to obtain target service data associated with a target object, and determine a matching feature between the target object and the target service data;
wherein the first determining module 11 comprises: a data acquisition unit 111, a feature acquisition unit 112, a feature embedding unit 113, a vector determination unit 114, a dot product operation unit 115;
a data acquisition unit 111 for acquiring a service data set associated with a target object, and acquiring target service data from the service data set;
a feature acquiring unit 112, configured to acquire a pre-training object feature associated with a target object and a pre-training service feature associated with a service data sequence; the business data sequence is associated with the target object;
the service data sequence comprises a service trigger sequence and a service conversion sequence; the service trigger sequence comprises service data of the target object with trigger behavior, and the service conversion sequence comprises service data of the target object with conversion behavior;
A feature acquisition unit 112, specifically configured to input the object attribute of the target object, the service trigger sequence associated with the target object, and the service conversion sequence associated with the target object into the feature recognition network model; the feature recognition network model comprises a pre-training embedded layer, an attention processing layer and a pre-training connecting layer;
the feature obtaining unit 112 is specifically configured to perform feature embedding on the object attribute, the service data in the service trigger sequence, and the service data in the service conversion sequence through the pre-training embedding layer, so as to obtain an object feature corresponding to the object attribute, a trigger service feature corresponding to the service data in the service trigger sequence, and a conversion service feature corresponding to the service data in the service conversion sequence;
the feature obtaining unit 112 is specifically configured to input the object feature, the trigger service feature, and the conversion service feature to the attention processing layer, and perform attention processing on the object feature, the trigger service feature, and the conversion service feature through the attention processing layer, so as to obtain an attention object feature corresponding to the object feature, an attention trigger feature corresponding to the trigger service feature, and an attention conversion feature corresponding to the conversion service feature;
The feature obtaining unit 112 is specifically configured to input the object feature, the attention triggering feature, and the attention conversion feature to a pre-training connection layer, perform feature fusion on the object feature and the attention object feature through the pre-training connection layer, obtain a pre-training object feature associated with the target object, and generate, in the pre-training connection layer, a pre-training service feature corresponding to service data included in the service triggering sequence and the service conversion sequence according to the attention triggering feature and the attention conversion feature.
The feature embedding unit 113 is configured to input an object attribute of the target object, a service attribute of the target service data, a pre-training object feature, and a pre-training service feature into a matching sub-network in the probability estimation network model; the matching sub-network comprises a matching embedding layer;
the feature embedding unit 113 is configured to perform feature embedding on the object attribute and the service attribute through the matching embedding layer, so as to obtain an object feature corresponding to the object attribute and a service feature corresponding to the service attribute;
a vector determining unit 114, configured to determine, in the matching sub-network, an object feature vector associated with the target object according to the object feature and the pre-training object feature, and determine a service feature vector associated with the target service data according to the service feature and the pre-training service feature;
The matching sub-network comprises an object full-connection layer, a service full-connection layer and a characteristic average layer; the service data sequence comprises one or more service data, and each service data corresponds to a pre-training service feature respectively;
the vector determining unit 114 is specifically configured to input the object feature to the object full-connection layer in the matching sub-network, and perform full-connection processing on the object feature through the object full-connection layer to obtain a full-connection object feature;
the vector determining unit 114 is specifically configured to perform feature fusion on the fully connected object feature and the pre-trained object feature to obtain an object feature vector associated with the target object;
the vector determining unit 114 is specifically configured to input the service feature to a service full-connection layer, and perform full-connection processing on the service feature through the service full-connection layer to obtain a full-connection service feature;
the vector determining unit 114 is specifically configured to input one or more pre-training service features to a feature averaging layer, and average the one or more pre-training service features through the feature averaging layer to obtain an average pre-training service feature;
the vector determining unit 114 is specifically configured to perform feature fusion on the full-connection service feature and the average pre-training service feature, so as to obtain a service feature vector associated with the target service data.
The dot product operation unit 115 is configured to perform dot product operation on the object feature vector and the service feature vector, so as to obtain a matching feature between the target object and the target service data.
The specific implementation manners of the data acquisition unit 111, the feature acquisition unit 112, the feature embedding unit 113, the vector determination unit 114 and the dot product operation unit 115 may be referred to the description of step S101 and step S1011-step S1019 in the embodiment corresponding to fig. 4 in the embodiment corresponding to fig. 3, and will not be repeated here.
A second determining module 12, configured to determine a target service triggering characteristic and a target service conversion characteristic of the target object for the target service data; the target service conversion feature is obtained by carrying out feature sharing transfer processing on the target service triggering feature;
wherein the probability estimation network model further comprises a sequencing subnetwork; the sequencing sub-network comprises an input network layer, a parameter sharing network layer and a multi-layer perception network layer;
the second determination module 12 includes: a feature determination unit 121, a feature generation unit 122, a feature transfer unit 123;
a feature determining unit 121, configured to determine, through an input network layer, a shared attribute feature of a target object for target service data;
The input network layer comprises a characteristic embedding layer and a characteristic splicing layer; the service data sequence comprises one or more service data, and each service data corresponds to a pre-training service feature respectively;
the feature determining unit 121 is specifically configured to input an object attribute of the target object, a service attribute of the target service data, and a context attribute associated with the target object to the feature embedding layer, and perform feature embedding on the object attribute, the service attribute, and the context attribute through the feature embedding layer to obtain an object feature corresponding to the object attribute, a service feature corresponding to the service attribute, and a context feature corresponding to the context attribute;
the feature determining unit 121 is specifically configured to perform average processing on one or more pre-training service features to obtain an average pre-training service feature;
the feature determining unit 121 is specifically configured to input the object feature, the pre-training object feature, the service feature, and the average pre-training service feature into a feature stitching layer, perform feature stitching on the object feature and the pre-training object feature in the feature stitching layer to obtain a stitched object feature, and perform feature stitching on the service feature and the average pre-training service feature to obtain a stitched service feature;
The feature determining unit 121 is specifically configured to perform feature stitching on the stitching object feature, the stitching service feature, and the context feature, so as to obtain a shared attribute feature of the target object for the target service data.
The feature generating unit 122 is configured to input the shared attribute feature to the parameter sharing network layer, and generate an initial service triggering feature and an initial service conversion feature of the target object for the target service data through the parameter sharing network layer;
the parameter sharing network layer comprises a sharing full-connection layer, a weight learning layer, a feature classification layer, a first triggering full-connection layer, a first shallow conversion full-connection layer and a first deep conversion full-connection layer; the initial business transformation features comprise initial shallow transformation features and initial deep transformation features;
the feature generating unit 122 is specifically configured to input the shared attribute feature to a shared full-connection layer, and perform full-connection processing on the shared attribute feature through the shared full-connection layer to obtain a full-connection shared feature;
the feature generating unit 122 is specifically configured to input the full-connection sharing feature to the weight learning layer, and perform feature weighting processing on the full-connection sharing feature through the weight learning layer to obtain a weight sharing feature;
The feature generation unit 122 is specifically configured to input the weight sharing feature to a feature classification layer, and perform feature classification on the weight sharing feature by using the feature classification layer to obtain a trigger distribution vector, a shallow conversion distribution vector, and a deep conversion distribution vector of the target object for the target service data;
the feature generating unit 122 is specifically configured to input a trigger distribution vector to a first trigger full-connection layer, perform full-connection processing on the trigger distribution vector through the first trigger full-connection layer, and generate an initial service trigger feature of the target object for the target service data;
the feature generation unit 122 is specifically configured to input a shallow conversion distribution vector to a first shallow conversion full-connection layer, perform full-connection processing on the shallow conversion distribution vector through the first shallow conversion full-connection layer, and generate an initial shallow conversion feature of the target object for the target service data;
the feature generation unit 122 is specifically configured to input the deep conversion distribution vector to a first deep conversion full-connection layer, and perform full-connection processing on the deep conversion distribution vector through the first deep conversion full-connection layer, so as to generate an initial deep conversion feature of the target object for the target service data.
The weight learning layer comprises weight learning components corresponding to object features, business features, context features, pre-training object features and average pre-training business features respectively; the feature classification layer comprises a triggering feature classification layer, a shallow conversion feature classification layer and a deep conversion feature classification layer;
the feature generating unit 122 is specifically configured to input the full-connection shared feature to each weight learning component, and perform feature weighting processing on the full-connection shared feature through each weight learning component to obtain a weight shared feature corresponding to each weight learning component;
the feature generation unit 122 is specifically configured to perform feature fusion on the weight sharing features corresponding to each weight learning component, so as to obtain fusion sharing features;
the feature generation unit 122 is specifically configured to input the fusion shared feature to a trigger feature classification layer, and perform feature classification on the fusion shared feature through the trigger feature classification layer to obtain a trigger distribution vector of the target object for the target service data;
the feature generation unit 122 is specifically configured to input the fusion shared feature to a shallow transformation feature classification layer, and perform feature classification on the fusion shared feature through the shallow transformation feature classification layer to obtain a shallow transformation distribution vector of the target object for the target service data;
The feature generation unit 122 is specifically configured to input the fusion shared feature to a deep transformation feature classification layer, and perform feature classification on the fusion shared feature by the deep transformation feature classification layer to obtain a deep transformation distribution vector of the target object for the target service data.
The feature transfer unit 123 is configured to input the initial service triggering feature and the initial service conversion feature into a multi-layer sensing network layer, perform full connection processing on the initial service triggering feature through the multi-layer sensing network layer to obtain a target service triggering feature of the target object for the target service data, and perform feature sharing transfer processing on the target service triggering feature and the initial service conversion feature in the multi-layer sensing network layer to obtain a target service conversion feature of the target object for the target service data.
The multi-layer sensing network layer comprises a second triggering full-connection layer, a second shallow conversion full-connection layer, a shallow connection layer, a second deep conversion full-connection layer and a deep connection layer; the initial business transformation features comprise initial shallow transformation features and initial deep transformation features; the target business transformation features comprise target shallow transformation features and target deep transformation features;
The feature transfer unit 123 is specifically configured to input an initial service triggering feature to the second triggering full-connection layer, and perform full-connection processing on the initial service triggering feature through the second triggering full-connection layer to obtain a target service triggering feature of the target object for target service data;
the feature transfer unit 123 is specifically configured to input the target service triggering feature and the initial shallow conversion feature to a shallow connection layer, and perform feature stitching on the target service triggering feature and the initial shallow conversion feature through the shallow connection layer to obtain a shared shallow stitching feature;
the feature transfer unit 123 is specifically configured to input the shared shallow splicing feature to a second shallow conversion full-connection layer, and perform full-connection processing on the shared shallow splicing feature through the second shallow conversion full-connection layer to obtain a target shallow conversion feature of the target object for the target service data;
the feature transfer unit 123 is specifically configured to input the target shallow layer transformation feature and the initial deep layer transformation feature into the deep layer, and perform feature stitching on the target shallow layer transformation feature and the initial deep layer transformation feature through the deep layer connection layer to obtain a shared deep layer stitching feature;
the feature transfer unit 123 is specifically configured to input the shared deep splicing feature to a second deep conversion full-connection layer, and perform full-connection processing on the shared deep splicing feature through the second deep conversion full-connection layer, so as to obtain a target deep conversion feature of the target object for the target service data.
For specific implementation manners of the feature determining unit 121, the feature generating unit 122 and the feature transferring unit 123, reference may be made to the description of step S102 and step S1023 in the embodiment corresponding to fig. 3 and the embodiment corresponding to fig. 6, and the description will not be repeated here.
The first estimating module 13 is configured to perform a stitching process on the matching feature and the target service triggering feature, obtain an initial stitching triggering feature of the target object for the target service data, and determine a triggering probability estimated value of the target object for the target service data based on the initial stitching triggering feature;
the ordering sub-network further comprises a characteristic connecting layer and a dot product full connecting layer;
the first estimation module 13 includes: the first splicing unit 131, the first processing unit 132, and the first estimating unit 133;
the first splicing unit 131 is configured to input the matching feature and the target service triggering feature to the feature connection layer, and splice the matching feature and the target service triggering feature through the feature connection layer to obtain an initial splicing triggering feature of the target object for the target service data;
the first processing unit 132 is configured to input the initial splice trigger feature to the dot product full-connection layer, and perform full-connection processing on the initial splice trigger feature through the dot product full-connection layer to obtain a target splice trigger feature;
The first estimating unit 133 is configured to determine a trigger probability estimated value of the target object for the target service data according to the target splicing trigger feature.
The specific implementation manner of the first splicing unit 131, the first processing unit 132, and the first estimating unit 133 may refer to the description of step S103 in the embodiment corresponding to fig. 3, and will not be repeated here.
The second estimating module 14 is configured to perform a stitching process on the matching feature and the target service transformation feature, obtain an initial stitching transformation feature of the target object for the target service data, and determine a transformation probability estimated value of the target object for the target service data based on the initial stitching transformation feature.
The ordering sub-network further comprises a characteristic connecting layer and a dot product full connecting layer; the conversion probability estimated value comprises a shallow conversion probability estimated value and a deep conversion probability estimated value; the target business transformation features comprise target shallow transformation features and target deep transformation features;
the second estimation module 14 includes: the system comprises a second splicing unit 141, a second processing unit 142, a second estimating unit 143, a third splicing unit 144, a third processing unit 145 and a third estimating unit 146;
the second splicing unit 141 is configured to input the matching feature and the target shallow conversion feature to the feature connection layer, and splice the matching feature and the target shallow conversion feature through the feature connection layer to obtain an initial spliced shallow conversion feature of the target object for the target service data;
The second processing unit 142 is configured to input the initial stitching shallow layer conversion feature to the dot product full-connection layer, and perform full-connection processing on the initial stitching shallow layer conversion feature through the dot product full-connection layer to obtain a target stitching shallow layer conversion feature;
the second estimating unit 143 is configured to determine a shallow conversion probability estimated value of the target object for the target service data according to the target spliced shallow conversion characteristic;
the third stitching unit 144 is configured to input the matching feature and the target deep conversion feature to the feature connection layer, and perform stitching processing on the matching feature and the target deep conversion feature through the feature connection layer to obtain an initial stitched deep conversion feature of the target object for the target service data;
the third processing unit 145 is configured to input the initial stitching deep layer transformation characteristic to the dot product full-connection layer, and perform full-connection processing on the initial stitching deep layer transformation characteristic through the dot product full-connection layer to obtain a target stitching deep layer transformation characteristic;
the third estimating unit 146 is configured to determine a deep transformation probability estimated value of the target object for the target service data according to the target stitching deep transformation characteristic.
The specific implementation manner of the second splicing unit 141, the second processing unit 142, the second estimating unit 143, the third splicing unit 144, the third processing unit 145 and the third estimating unit 146 may be referred to the description of step S104 in the embodiment corresponding to fig. 3, and will not be repeated here.
The specific implementation manners of the first determining module 11, the second determining module 12, the first estimating module 13 and the second estimating module 14 may be referred to in the embodiment corresponding to fig. 3, and the description of the step S101 to the step S104, the step S1011 to the step S1019 in the embodiment corresponding to fig. 4, the step S1021 to the step S1023 in the embodiment corresponding to fig. 6, and the step S201 to the step S207 in the embodiment corresponding to fig. 7 will not be repeated here. In addition, the description of the beneficial effects of the same method is omitted.
Further, referring to fig. 12, fig. 12 is a schematic structural diagram of a data processing apparatus according to an embodiment of the present application. The data processing apparatus 2 may include: a first sample determination module 21, a second sample determination module 22, a first sample estimation module 23, a second sample estimation module 24, and a parameter adjustment module 25;
a first sample determining module 21, configured to obtain sample service data associated with a sample object, and determine a sample matching feature between the sample object and the sample service data through an initial probability pre-estimation network model;
a second sample determining module 22, configured to determine, in the initial probability estimation network model, a sample service triggering characteristic and a sample service conversion characteristic of the sample object for sample service data; the sample service conversion feature is obtained by carrying out feature sharing transfer processing on the sample service triggering feature;
The first sample estimating module 23 is configured to perform a stitching process on the sample matching feature and the sample service triggering feature to obtain a sample stitching triggering feature of the sample object for the sample service data, and determine a sample triggering probability pre-estimated value of the sample object for the sample service data based on the sample stitching triggering feature;
the second sample estimating module 24 is configured to perform a stitching process on the sample matching feature and the sample service conversion feature to obtain a sample stitching conversion feature of the sample object for the sample service data, and determine a sample conversion probability estimated value of the sample object for the sample service data based on the sample stitching conversion feature;
the parameter adjustment module 25 is configured to perform parameter adjustment on the initial probability prediction network model based on sample tag information, a sample trigger probability prediction value and a sample conversion probability prediction value of the sample object for the sample service data, and take the initial probability prediction network model after parameter adjustment as a probability prediction network model; the probability prediction network model is used for predicting a trigger probability predicted value and a conversion probability predicted value of the target object aiming at the target service data.
The specific implementation manner of the first sample determining module 21, the second sample determining module 22, the first sample estimating module 23, the second sample estimating module 24 and the parameter adjusting module 25 may be referred to the description of step S301 to step S305 in the embodiment corresponding to fig. 10, and will not be repeated here. In addition, the description of the beneficial effects of the same method is omitted.
Further, referring to fig. 13, fig. 13 is a schematic structural diagram of a computer device according to an embodiment of the present application. As shown in fig. 13, the computer device 1000 may include: processor 1001, network interface 1004, and memory 1005, and in addition, the above-described computer device 1000 may further include: a user interface 1003, and at least one communication bus 1002. Wherein the communication bus 1002 is used to enable connected communication between these components. In some embodiments, the user interface 1003 may include a Display (Display), a Keyboard (Keyboard), and the optional user interface 1003 may further include a standard wired interface, a wireless interface, among others. Alternatively, the network interface 1004 may include a standard wired interface, a wireless interface (e.g., WI-FI interface). The memory 1005 may be a high-speed RAM memory or a non-volatile memory (non-volatile memory), such as at least one disk memory. Optionally, the memory 1005 may also be at least one memory device located remotely from the aforementioned processor 1001. As shown in fig. 13, an operating system, a network communication module, a user interface module, and a device control application program may be included in the memory 1005, which is one type of computer-readable storage medium.
In the computer device 1000 shown in FIG. 13, the network interface 1004 may provide network communication functions; while user interface 1003 is primarily used as an interface for providing input to a user; and the processor 1001 may be used to invoke device control applications stored in the memory 1005.
It should be understood that the computer device 1000 described in the embodiments of the present application may perform the description of the data processing method in the embodiments corresponding to fig. 3, 4, 6, 7 and 10, and may also perform the description of the data processing device 1 in the embodiments corresponding to fig. 11 and the description of the data processing device 2 in the embodiments corresponding to fig. 12, which are not repeated herein. In addition, the description of the beneficial effects of the same method is omitted.
Furthermore, it should be noted here that: the embodiments of the present application further provide a computer readable storage medium, in which the aforementioned computer program executed by the data processing apparatus 1 and the data processing apparatus 2 is stored, and the computer program includes program instructions, when executed by a processor, can execute the description of the data processing method in the embodiments corresponding to fig. 3, 4, 6, 7 and 10, and therefore, will not be described in detail herein. In addition, the description of the beneficial effects of the same method is omitted. For technical details not disclosed in the embodiments of the computer-readable storage medium according to the present application, please refer to the description of the method embodiments of the present application.
In addition, it should be noted that: embodiments of the present application also provide a computer program product or computer program that may include computer instructions that may be stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor may execute the computer instructions, so that the computer device performs the description of the data processing method in the embodiments corresponding to fig. 3, fig. 4, fig. 6, fig. 7, and fig. 10, which will not be described herein. In addition, the description of the beneficial effects of the same method is omitted. For technical details not disclosed in the computer program product or the computer program embodiments related to the present application, please refer to the description of the method embodiments of the present application.
Those skilled in the art will appreciate that implementing all or part of the above-described methods may be accomplished by way of a computer program stored in a computer-readable storage medium, which when executed may comprise the steps of the embodiments of the methods described above. The storage medium may be a magnetic disk, an optical disk, a Read-Only Memory (ROM), a random access Memory (Random Access Memory, RAM), or the like.
The foregoing disclosure is only illustrative of the preferred embodiments of the present application and is not intended to limit the scope of the claims herein, as the equivalent of the claims herein shall be construed to fall within the scope of the claims herein.
Claims (15)
1. A method of data processing, comprising:
acquiring target business data associated with a target object, and determining matching characteristics between the target object and the target business data;
determining target service triggering characteristics and target service conversion characteristics of the target object aiming at the target service data; the target service conversion feature is obtained by carrying out feature sharing transfer processing on the target service triggering feature;
performing splicing processing on the matching features and the target service triggering features to obtain initial splicing triggering features of the target object aiming at the target service data, and determining triggering probability pre-estimated values of the target object aiming at the target service data based on the initial splicing triggering features;
and performing splicing processing on the matching characteristic and the target service conversion characteristic to obtain an initial splicing conversion characteristic of the target object aiming at the target service data, and determining a conversion probability estimated value of the target object aiming at the target service data based on the initial splicing conversion characteristic.
2. The method of claim 1, wherein the obtaining the target business data associated with the target object, determining the matching characteristics between the target object and the target business data, comprises:
acquiring a service data set associated with a target object, and acquiring target service data from the service data set;
acquiring pre-training object features associated with the target object and pre-training service features associated with a service data sequence; the business data sequence is associated with the target object;
inputting the object attribute of the target object, the service attribute of the target service data, the pre-training object feature and the pre-training service feature into a matching sub-network in a probability estimation network model; the matching sub-network comprises a matching embedding layer;
performing feature embedding on the object attribute and the service attribute through the matching embedding layer to obtain an object feature corresponding to the object attribute and a service feature corresponding to the service attribute;
in the matching sub-network, determining an object feature vector associated with the target object according to the object feature and the pre-training object feature, and determining a service feature vector associated with the target service data according to the service feature and the pre-training service feature;
And performing dot product operation on the object feature vector and the service feature vector to obtain the matching feature between the target object and the target service data.
3. The method of claim 2, wherein the service data sequence comprises a service trigger sequence and a service conversion sequence; the service triggering sequence comprises service data of the target object with triggering behavior, and the service conversion sequence comprises service data of the target object with conversion behavior;
the acquiring pre-training object features associated with the target object and pre-training business features associated with a business data sequence comprises:
inputting the object attribute of the target object, a service trigger sequence associated with the target object, and a service conversion sequence associated with the target object into a feature recognition network model; the feature recognition network model comprises a pre-training embedded layer, an attention processing layer and a pre-training connecting layer;
the feature embedding is respectively carried out on the object attribute, the service data in the service trigger sequence and the service data in the service conversion sequence through the pre-training embedding layer, so as to obtain an object feature corresponding to the object attribute, a triggering service feature corresponding to the service data in the service trigger sequence and a conversion service feature corresponding to the service data in the service conversion sequence;
Inputting the object feature, the triggering service feature and the conversion service feature to the attention processing layer, and performing attention processing on the object feature, the triggering service feature and the conversion service feature through the attention processing layer to obtain an attention object feature corresponding to the object feature, an attention triggering feature corresponding to the triggering service feature and an attention conversion feature corresponding to the conversion service feature;
inputting the object features, the attention triggering features and the attention conversion features to the pre-training connection layer, carrying out feature fusion on the object features and the attention object features through the pre-training connection layer to obtain pre-training object features associated with the target object, and generating pre-training service features corresponding to the service triggering sequence and the service data contained in the service conversion sequence according to the attention triggering features and the attention conversion features in the pre-training connection layer.
4. The method of claim 2, wherein the matching subnetwork comprises an object fully connected layer, a traffic fully connected layer, and a feature averaging layer; the service data sequence comprises one or more service data, and each service data corresponds to a pre-training service feature respectively;
In the matching sub-network, determining an object feature vector associated with the target object according to the object feature and the pre-training object feature, and determining a service feature vector associated with the target service data according to the service feature and the pre-training service feature, including:
in the matching sub-network, inputting the object features to the object full-connection layer, and performing full-connection processing on the object features through the object full-connection layer to obtain full-connection object features;
feature fusion is carried out on the fully connected object features and the pre-trained object features to obtain object feature vectors associated with the target objects;
inputting the service characteristics to the service full-connection layer, and carrying out full-connection processing on the service characteristics through the service full-connection layer to obtain full-connection service characteristics;
inputting one or more pre-training service features into a feature average layer, and carrying out average processing on the one or more pre-training service features through the feature average layer to obtain average pre-training service features;
and carrying out feature fusion on the full-connection service features and the average pre-training service features to obtain service feature vectors associated with the target service data.
5. The method of claim 2, wherein the probabilistic predictive network model further comprises a ranking sub-network; the sequencing sub-network comprises an input network layer, a parameter sharing network layer and a multi-layer perception network layer;
the determining the target service triggering characteristic and the target service conversion characteristic of the target object aiming at the target service data comprises the following steps:
determining a sharing attribute characteristic of the target object aiming at the target service data through the input network layer;
inputting the shared attribute characteristics to the parameter sharing network layer, and generating initial service triggering characteristics and initial service conversion characteristics of the target object aiming at the target service data through the parameter sharing network layer;
and inputting the initial service triggering characteristics and the initial service conversion characteristics into the multi-layer sensing network layer, performing full connection processing on the initial service triggering characteristics through the multi-layer sensing network layer to obtain target service triggering characteristics of the target object aiming at the target service data, and performing characteristic sharing transmission processing on the target service triggering characteristics and the initial service conversion characteristics in the multi-layer sensing network layer to obtain target service conversion characteristics of the target object aiming at the target service data.
6. The method of claim 5, wherein the input network layer comprises a feature embedding layer and a feature stitching layer; the service data sequence comprises one or more service data, and each service data corresponds to a pre-training service feature respectively;
the determining, by the input network layer, a shared attribute feature of the target object for the target service data includes:
inputting the object attribute of the target object, the service attribute of the target service data and the context attribute associated with the target object into the feature embedding layer, and performing feature embedding on the object attribute, the service attribute and the context attribute through the feature embedding layer to obtain an object feature corresponding to the object attribute, a service feature corresponding to the service attribute and a context feature corresponding to the context attribute;
carrying out average processing on one or more pre-training service characteristics to obtain average pre-training service characteristics;
inputting the object features, the pre-training object features, the service features and the average pre-training service features into the feature stitching layer, performing feature stitching on the object features and the pre-training object features in the feature stitching layer to obtain stitched object features, and performing feature stitching on the service features and the average pre-training service features to obtain stitched service features;
And performing feature stitching on the stitching object features, the stitching business features and the context features to obtain the sharing attribute features of the target object aiming at the target business data.
7. The method of claim 6, wherein the parameter sharing network layer comprises a sharing full connection layer, a weight learning layer, a feature classification layer, a first triggering full connection layer, a first shallow conversion full connection layer, and a first deep conversion full connection layer; the initial business transformation characteristics comprise initial shallow transformation characteristics and initial deep transformation characteristics;
the step of inputting the shared attribute feature to the parameter sharing network layer, generating an initial service triggering feature and an initial service conversion feature of the target object for the target service data through the parameter sharing network layer, including:
inputting the shared attribute features to the shared full-connection layer, and carrying out full-connection processing on the shared attribute features through the shared full-connection layer to obtain full-connection shared features;
inputting the full-connection sharing feature to the weight learning layer, and carrying out feature weighting processing on the full-connection sharing feature through the weight learning layer to obtain a weight sharing feature;
Inputting the weight sharing characteristics to the characteristic classification layer, and carrying out characteristic classification on the weight sharing characteristics through the characteristic classification layer to obtain a trigger distribution vector, a shallow conversion distribution vector and a deep conversion distribution vector of the target object aiming at the target service data;
inputting the trigger distribution vector to the first trigger full-connection layer, and performing full-connection processing on the trigger distribution vector through the first trigger full-connection layer to generate an initial service trigger characteristic of the target object aiming at the target service data;
inputting the shallow conversion distribution vector to the first shallow conversion full-connection layer, and performing full-connection processing on the shallow conversion distribution vector through the first shallow conversion full-connection layer to generate initial shallow conversion characteristics of the target object aiming at the target service data;
and inputting the deep conversion distribution vector to the first deep conversion full-connection layer, and performing full-connection processing on the deep conversion distribution vector through the first deep conversion full-connection layer to generate initial deep conversion characteristics of the target object aiming at the target business data.
8. The method of claim 7, wherein the weight learning layer comprises weight learning components corresponding to the object features, the business features, the context features, the pre-trained object features, and the average pre-trained business features, respectively; the feature classification layer comprises a triggering feature classification layer, a shallow conversion feature classification layer and a deep conversion feature classification layer;
the step of inputting the full-connection sharing feature to the weight learning layer, and performing feature weighting processing on the full-connection sharing feature by the weight learning layer to obtain a weight sharing feature comprises the following steps:
the full-connection sharing feature is respectively input to each weight learning component, and feature weighting processing is respectively carried out on the full-connection sharing feature through each weight learning component to obtain weight sharing features respectively corresponding to each weight learning component;
inputting the weight sharing feature to the feature classification layer, and performing feature classification on the weight sharing feature by the feature classification layer to obtain a trigger distribution vector, a shallow conversion distribution vector and a deep conversion distribution vector of the target object for the target service data, wherein the trigger distribution vector, the shallow conversion distribution vector and the deep conversion distribution vector comprise:
Feature fusion is carried out on the weight sharing features corresponding to each weight learning component respectively, so that fusion sharing features are obtained;
inputting the fusion shared features to the trigger feature classification layer, and performing feature classification on the fusion shared features through the trigger feature classification layer to obtain trigger distribution vectors of the target object aiming at the target service data;
inputting the fusion shared features to the shallow conversion feature classification layer, and carrying out feature classification on the fusion shared features through the shallow conversion feature classification layer to obtain shallow conversion distribution vectors of the target object aiming at the target service data;
and inputting the fusion shared features to the deep conversion feature classification layer, and carrying out feature classification on the fusion shared features through the deep conversion feature classification layer to obtain deep conversion distribution vectors of the target object aiming at the target service data.
9. The method of claim 5, wherein the multi-layer sensory network layer comprises a second trigger fully connected layer, a second shallow conversion fully connected layer, a shallow connected layer, a second deep conversion fully connected layer, and a deep connected layer; the initial business transformation characteristics comprise initial shallow transformation characteristics and initial deep transformation characteristics; the target business transformation characteristics comprise target shallow transformation characteristics and target deep transformation characteristics;
The step of inputting the initial service triggering characteristic and the initial service conversion characteristic to the multi-layer sensing network layer, performing full connection processing on the initial service triggering characteristic through the multi-layer sensing network layer to obtain a target service triggering characteristic of the target object aiming at the target service data, and performing characteristic sharing transfer processing on the target service triggering characteristic and the initial service conversion characteristic in the multi-layer sensing network layer to obtain a target service conversion characteristic of the target object aiming at the target service data, wherein the method comprises the steps of:
inputting the initial service triggering characteristic to the second triggering full-connection layer, and performing full-connection processing on the initial service triggering characteristic through the second triggering full-connection layer to obtain a target service triggering characteristic of the target object aiming at the target service data;
inputting the target service triggering characteristic and the initial shallow conversion characteristic to the shallow connecting layer, and performing characteristic splicing on the target service triggering characteristic and the initial shallow conversion characteristic through the shallow connecting layer to obtain a shared shallow splicing characteristic;
inputting the shared shallow splicing characteristic to the second shallow conversion full-connection layer, and carrying out full-connection processing on the shared shallow splicing characteristic through the second shallow conversion full-connection layer to obtain the target shallow conversion characteristic of the target object aiming at the target service data;
Inputting the target shallow layer transformation characteristic and the initial deep layer transformation characteristic into the deep connecting layer, and performing characteristic splicing on the target shallow layer transformation characteristic and the initial deep layer transformation characteristic through the deep connecting layer to obtain a shared deep splicing characteristic;
and inputting the shared deep splicing characteristic to the second deep conversion full-connection layer, and carrying out full-connection processing on the shared deep splicing characteristic through the second deep conversion full-connection layer to obtain the target deep conversion characteristic of the target object aiming at the target business data.
10. The method of claim 5, wherein the ordering subnetwork further comprises a feature connectivity layer and a dot product full connectivity layer;
the splicing processing is performed on the matching feature and the target service triggering feature to obtain an initial splicing triggering feature of the target object for the target service data, and the determining of the triggering probability pre-estimated value of the target object for the target service data based on the initial splicing triggering feature comprises the following steps:
inputting the matching feature and the target service triggering feature into the feature connection layer, and performing splicing processing on the matching feature and the target service triggering feature through the feature connection layer to obtain an initial splicing triggering feature of the target object aiming at the target service data;
Inputting the initial splicing triggering characteristics to the dot product full-connection layer, and performing full-connection processing on the initial splicing triggering characteristics through the dot product full-connection layer to obtain target splicing triggering characteristics;
and determining a trigger probability estimated value of the target object aiming at the target service data according to the target splicing trigger characteristics.
11. The method of claim 5, wherein the ordering subnetwork further comprises a feature connectivity layer and a dot product full connectivity layer; the transformation probability estimated value comprises a shallow transformation probability estimated value and a deep transformation probability estimated value; the target business transformation characteristics comprise target shallow transformation characteristics and target deep transformation characteristics;
the splicing processing is performed on the matching feature and the target service conversion feature to obtain an initial splicing conversion feature of the target object for the target service data, and the determining a conversion probability pre-estimation value of the target object for the target service data based on the initial splicing conversion feature comprises the following steps:
inputting the matching feature and the target shallow conversion feature into the feature connection layer, and performing splicing treatment on the matching feature and the target shallow conversion feature through the feature connection layer to obtain an initial spliced shallow conversion feature of the target object aiming at the target service data;
Inputting the initial splicing shallow layer conversion characteristics to the dot product full-connection layer, and carrying out full-connection processing on the initial splicing shallow layer conversion characteristics through the dot product full-connection layer to obtain target splicing shallow layer conversion characteristics;
determining shallow conversion probability pre-estimation values of the target object aiming at the target service data according to the target splicing shallow conversion characteristics;
inputting the matching feature and the target deep conversion feature into the feature connection layer, and performing splicing treatment on the matching feature and the target deep conversion feature through the feature connection layer to obtain an initial spliced deep conversion feature of the target object aiming at the target service data;
inputting the initial splicing deep conversion characteristics to the dot product full-connection layer, and carrying out full-connection processing on the initial splicing deep conversion characteristics through the dot product full-connection layer to obtain target splicing deep conversion characteristics;
and determining a deep conversion probability estimated value of the target object aiming at the target service data according to the target splicing deep conversion characteristics.
12. A method of data processing, comprising:
acquiring sample service data associated with a sample object, and determining sample matching characteristics between the sample object and the sample service data through an initial probability pre-estimated network model;
In the initial probability pre-estimated network model, determining sample service triggering characteristics and sample service conversion characteristics of the sample object aiming at the sample service data; the sample service conversion characteristics are obtained by carrying out characteristic sharing transfer processing on the sample service triggering characteristics;
performing splicing processing on the sample matching characteristics and the sample service triggering characteristics to obtain sample splicing triggering characteristics of the sample object aiming at the sample service data, and determining a sample triggering probability pre-estimated value of the sample object aiming at the sample service data based on the sample splicing triggering characteristics;
performing splicing processing on the sample matching characteristics and the sample service conversion characteristics to obtain sample splicing conversion characteristics of the sample object aiming at the sample service data, and determining a sample conversion probability estimated value of the sample object aiming at the sample service data based on the sample splicing conversion characteristics;
based on sample label information, the sample triggering probability estimated value and the sample conversion probability estimated value of the sample object aiming at the sample service data, carrying out parameter adjustment on the initial probability estimated network model, and taking the initial probability estimated network model after parameter adjustment as a probability estimated network model; the probability prediction network model is used for predicting a trigger probability predicted value and a conversion probability predicted value of the target object aiming at the target service data.
13. A computer device, comprising: a processor and a memory;
the processor is connected to the memory, wherein the memory is configured to store a computer program, and the processor is configured to invoke the computer program to cause the computer device to perform the method of any of claims 1-12.
14. A computer readable storage medium, characterized in that the computer readable storage medium has stored therein a computer program adapted to be loaded and executed by a processor to cause a computer device having the processor to perform the method of any of claims 1-12.
15. A computer program product, characterized in that it comprises computer instructions stored in a computer-readable storage medium and adapted to be read and executed by a processor to cause a computer device with the processor to perform the method of any of claims 1-12.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111385324.8A CN116167798A (en) | 2021-11-22 | 2021-11-22 | Data processing method, computer equipment and readable storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111385324.8A CN116167798A (en) | 2021-11-22 | 2021-11-22 | Data processing method, computer equipment and readable storage medium |
Publications (1)
Publication Number | Publication Date |
---|---|
CN116167798A true CN116167798A (en) | 2023-05-26 |
Family
ID=86411794
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111385324.8A Pending CN116167798A (en) | 2021-11-22 | 2021-11-22 | Data processing method, computer equipment and readable storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116167798A (en) |
-
2021
- 2021-11-22 CN CN202111385324.8A patent/CN116167798A/en active Pending
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111444428B (en) | Information recommendation method and device based on artificial intelligence, electronic equipment and storage medium | |
US20220215032A1 (en) | Ai-based recommendation method and apparatus, electronic device, and storage medium | |
CN111538912B (en) | Content recommendation method, device, equipment and readable storage medium | |
CN115917535A (en) | Recommendation model training method, recommendation device and computer readable medium | |
CN111242748B (en) | Method, apparatus, and storage medium for recommending items to a user | |
CN110020022B (en) | Data processing method, device, equipment and readable storage medium | |
CN110597962A (en) | Search result display method, device, medium and electronic equipment | |
CN111552835A (en) | File recommendation method and device and server | |
CN116452263A (en) | Information recommendation method, device, equipment, storage medium and program product | |
CN112269943A (en) | Information recommendation system and method | |
CN116956183A (en) | Multimedia resource recommendation method, model training method, device and storage medium | |
CN111860870A (en) | Training method, device, equipment and medium for interactive behavior determination model | |
CN109299351B (en) | Content recommendation method and device, electronic equipment and computer readable medium | |
CN116956204A (en) | Network structure determining method, data predicting method and device of multi-task model | |
CN116975426A (en) | Service data processing method, device, equipment and medium | |
CN114298118B (en) | Data processing method based on deep learning, related equipment and storage medium | |
CN114328992A (en) | Multimedia information recommendation method, device, program product, equipment and medium | |
CN115329183A (en) | Data processing method, device, storage medium and equipment | |
CN116167798A (en) | Data processing method, computer equipment and readable storage medium | |
CN116166870A (en) | Data processing method, computer equipment and readable storage medium | |
CN114428811B (en) | Data processing method and device, computer equipment and readable storage medium | |
CN116628236B (en) | Method and device for delivering multimedia information, electronic equipment and storage medium | |
CN114417875B (en) | Data processing method, apparatus, device, readable storage medium, and program product | |
CN114417944B (en) | Recognition model training method and device, and user abnormal behavior recognition method and device | |
CN114880565B (en) | Information recommendation method, model training method and related device |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
REG | Reference to a national code |
Ref country code: HK Ref legal event code: DE Ref document number: 40086772 Country of ref document: HK |
|
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |