CN111598599A - User characterization method and device, electronic equipment and computer readable medium - Google Patents

User characterization method and device, electronic equipment and computer readable medium Download PDF

Info

Publication number
CN111598599A
CN111598599A CN201910131144.3A CN201910131144A CN111598599A CN 111598599 A CN111598599 A CN 111598599A CN 201910131144 A CN201910131144 A CN 201910131144A CN 111598599 A CN111598599 A CN 111598599A
Authority
CN
China
Prior art keywords
data
model
domain
feature characterization
feature
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.)
Granted
Application number
CN201910131144.3A
Other languages
Chinese (zh)
Other versions
CN111598599B (en
Inventor
张心梓
陈海凯
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Alibaba Group Holding Ltd
Original Assignee
Alibaba Group Holding Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Alibaba Group Holding Ltd filed Critical Alibaba Group Holding Ltd
Priority to CN201910131144.3A priority Critical patent/CN111598599B/en
Publication of CN111598599A publication Critical patent/CN111598599A/en
Application granted granted Critical
Publication of CN111598599B publication Critical patent/CN111598599B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06Q30/0201—Market modelling; Market analysis; Collecting market data
    • 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Strategic Management (AREA)
  • Development Economics (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Image Analysis (AREA)
  • Computational Linguistics (AREA)
  • Marketing (AREA)
  • Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • General Business, Economics & Management (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)

Abstract

The embodiment of the application provides a user characterization method and device, electronic equipment and a computer readable medium, and relates to the field of artificial intelligence. Wherein the method comprises the following steps: training a cross-domain feature characterization model to be trained based on first feature data of a sample user in a source domain and second feature data of the sample user in a target domain; and generating first cross-domain feature characterization data of the user to be characterized in the target domain based on third feature data of the user to be characterized in the source domain through the trained cross-domain feature characterization model. According to the method and the device, the personalized characterization effect of the user with the missing target field characteristics in the target field can be effectively improved.

Description

User characterization method and device, electronic equipment and computer readable medium
Technical Field
The embodiment of the application relates to the field of artificial intelligence, in particular to a user characterization method, a user characterization device, electronic equipment and a computer readable medium.
Background
Commodity recommendation and commodity search related to cold-start users are important means for improving user liveness and pulling user quantity to increase in modern electronic commerce and are also important references of users for first impression of e-commerce platforms. However, the cold-start user usually lacks historical behaviors such as clicking or purchasing commodities of the e-commerce platform, the user portrait of the e-commerce platform is very sparse, and the cold-start user preference is estimated by using a traditional machine learning model, so that high accuracy is difficult to achieve. Due to the development of a large data platform, in many e-commerce companies, there are other fields besides the field of the commodity recommendation task, for example, treasure payment, high delicacy, pig flying, etc., and internet news, etc., besides the strict selection of internet. The account numbers of the two different fields are aligned, so that the characteristic information of the cold-start user in other fields can be acquired, and the personalized effect of the cold-start user on the commodity recommendation task in the target field can be greatly improved, so that cross-field commodity recommendation can be performed. The field in which feature information of the cold start user is rich is called a source field, and the field in which the commodity recommendation task is located is called a target field.
In a processing mode of a cross-domain commodity recommendation task, the traditional method is used for cascading feature information of a cold start user in a source domain and a target domain, inputting the cascaded feature information as the feature of the cold start user in the target domain, and learning the overall characterization of the cold start user. Specifically, the overall characterization of the cold-start user is learned by constructing a machine learning model for personalized commodity recommendation. However, the feature information of the target field is closer to the commodity recommendation task of the target field, and has a stronger characterization capability, the learning process inclines to the feature learning of the target field, the feature learning of the source field is insufficient, and the personalized characterization effect of the cold start user in the target field cannot be improved for the cold start user with serious feature loss in the target field.
Disclosure of Invention
The present application aims to provide a user characterization method, an apparatus, an electronic device, and a computer-readable medium, which are used to solve the problem in the prior art of how to effectively improve the personalized characterization effect of a user with missing features in a target field in the target field.
According to a first aspect of embodiments of the present application, a user characterization method is provided. The method comprises the following steps: training a cross-domain feature characterization model to be trained based on first feature data of a sample user in a source domain and second feature data of the sample user in a target domain; and generating first cross-domain feature characterization data of the user to be characterized in the target domain based on third feature data of the user to be characterized in the source domain through the trained cross-domain feature characterization model.
According to a second aspect of embodiments of the present application, there is provided a user characterization apparatus. The device comprises: the system comprises a first training module, a second training module and a third training module, wherein the first training module is used for training a cross-domain feature characterization model to be trained based on first feature data of a sample user in a source domain and second feature data of the sample user in a target domain; the first generation module is used for generating first cross-domain feature characterization data of the user to be characterized in the target domain based on third feature data of the user to be characterized in the source domain through the trained cross-domain feature characterization model.
According to a third aspect of embodiments of the present application, there is provided an electronic apparatus, including: one or more processors; a computer readable medium configured to store one or more programs which, when executed by the one or more processors, cause the one or more processors to implement the user characterization method as described in the first aspect of the embodiments above.
According to a fourth aspect of embodiments of the present application, there is provided a computer-readable medium, on which a computer program is stored, which when executed by a processor, implements the user characterization method as described in the first aspect of the embodiments above.
According to the technical scheme provided by the embodiment of the application, the cross-domain feature characterization model to be trained is trained based on the first feature data of the sample user in the source domain and the second feature data of the sample user in the target domain, the first cross-domain feature characterization data of the user to be characterized in the target domain is generated based on the third feature data of the user to be characterized in the source domain through the trained cross-domain feature characterization model, compared with the existing other modes, the cross-domain feature characterization data of the user to be characterized in the target domain is generated based on the feature data of the user to be characterized in the source domain, and the personalized characterization effect of the user with different degrees of missing features in the target domain can be effectively improved.
Drawings
Other features, objects and advantages of the present application will become more apparent upon reading of the following detailed description of non-limiting embodiments thereof, made with reference to the accompanying drawings in which:
FIG. 1 is a flowchart illustrating steps of a user characterization method according to an embodiment of the present disclosure;
FIG. 2 is a schematic diagram of a training process of a cross-domain feature characterization model according to an embodiment of the present application;
FIG. 3 is a flowchart illustrating steps of a user characterization method according to a second embodiment of the present application;
FIG. 4 is a diagram illustrating a framework for processing an object pre-estimation task according to a second embodiment of the present disclosure;
FIG. 5 is a schematic structural diagram of a user characterization device according to a third embodiment of the present application;
FIG. 6 is a schematic structural diagram of a user characterization device according to a fourth embodiment of the present application;
FIG. 7 is a schematic structural diagram of a user characterization device according to a fifth embodiment of the present application;
fig. 8 is a schematic structural diagram of an electronic device in a sixth embodiment of the present application;
fig. 9 is a hardware structure of an electronic device according to a seventh embodiment of the present application.
Detailed Description
The present application will be described in further detail with reference to the following drawings and examples. It is to be understood that the specific embodiments described herein are merely illustrative of the relevant invention and not restrictive of the invention. It should be noted that, for convenience of description, only the portions related to the related invention are shown in the drawings.
It should be noted that the embodiments and features of the embodiments in the present application may be combined with each other without conflict. The present application will be described in detail below with reference to the embodiments with reference to the attached drawings.
Referring to fig. 1, a flowchart illustrating steps of a user characterization method according to a first embodiment of the present application is shown.
Specifically, the user characterization method provided by this embodiment includes the following steps:
in step S101, a cross-domain feature characterization model to be trained is trained based on first feature data of a sample user in a source domain and second feature data of the sample user in a target domain.
In the embodiment of the present application, the sample user is an active user, that is, a user whose features in the source domain and those in the target domain are rich. The target field may be understood as a field where the current estimation task is located, for example, a commodity click rate estimation task or a commodity conversion rate estimation task is currently performed in the treasure panning client, and then the target field is the treasure panning client. The commodity click rate estimation task can be understood as a task for estimating the probability of a user browsing and clicking a commodity, and the commodity conversion rate estimation task can be understood as a task for estimating the probability of a user clicking and purchasing a commodity. For another example, if a novel click rate estimation task or a novel conversion rate estimation task is currently performed at the flag novel client, the target field is the flag novel client. The source domain may be understood as a domain having other features or knowledge available for migration to the target domain than the target domain. For example, when the target domain is a Taobao client, the source domain may be a Paobao client, a Goods client, a Fei pig client, and so on. For another example, when the target domain is a flag novel client, the source domain may be a treasure panning client, a treasure payment client, and the like. For another example, when the target domain is a cyber-easy-selected client, the source domain may be a cyber-easy news client. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In the embodiment of the present application, the first feature data may be understood as the overall feature data of the sample user in the source domain, including the behavior feature data and the user portrait data of the sample user in the source domain, for example, when the source domain is a pay client, a grand client, or a pig client, the first feature data may include the air ticket purchase data and the journey planning data of the sample user. The second feature data can be understood as the overall feature data of the sample user in the target field, including the behavior feature data and the user portrait data of the sample user in the target field, for example, when the target field is a panning client, the second feature data can include behavior sequence data of historical clicks and purchases of the sample user in the panning client, and purchasing power data, gender data, age data and the like of the sample user in the panning client. The cross-domain feature characterization model to be trained may be a neural network model composed of a plurality of fully connected layers, for example, a neural network model composed of three fully connected layers, a first fully connected layer may map input data to 1028-dimensional vectors, a second fully connected layer may map 1028-dimensional vectors to 512-dimensional vectors, and a third fully connected layer may map 512-dimensional vectors to 128-dimensional vectors. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In some optional embodiments, when a cross-domain feature characterization model to be trained is trained, generating, by the cross-domain feature characterization model, second cross-domain feature characterization data of the sample user in the target domain based on the first feature data; generating first target domain feature characterization data of the sample user based on the second feature data through the trained first target domain feature characterization model; training the cross-domain feature characterization model based on the second cross-domain feature characterization data and the first target domain feature characterization data. Therefore, the cross-domain feature characterization model can be trained simply and conveniently. It can be understood that any embodiment for training the cross-domain feature characterization model to be trained is applicable to this embodiment, and this is not limited in this embodiment of the present application.
In a specific example, before generating the first target domain feature characterization data of the sample user based on the second feature data by using the trained first target domain feature characterization model, the first target domain feature characterization model to be trained is trained based on the second feature data and the labeled target domain feature characterization data, so as to obtain the trained first target domain feature characterization model. Specifically, mapping operation is carried out on the second feature data through a first target domain feature characterization model to be trained, so as to obtain third target domain feature characterization data of the sample user; and training a first target field characteristic model to be trained based on the third target field characteristic data and the labeled target field characteristic data to obtain the trained first target field characteristic model. More specifically, determining a difference value between the third target domain feature characterization data and the labeled target domain feature characterization data through a target loss function; and adjusting parameters of the first target field feature characterization model based on the difference value. The training details of the first target domain feature representation model are similar to the training details of the cross-domain feature representation model described below, and are not repeated here. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In a specific example, when second cross-domain feature characterization data of the sample user in the target domain is generated, mapping operation is performed on the first feature data through the cross-domain feature characterization model to generate the second cross-domain feature characterization data of the sample user in the target domain. When the first target domain feature characterization data of the sample user is generated, mapping operation is performed on the second feature data through the trained first target domain feature characterization model, so that the first target domain feature characterization data of the sample user is generated. Determining a difference value between the second cross-domain feature characterization data and the first target domain feature characterization data through a target loss function when the cross-domain feature characterization model is trained based on the second cross-domain feature characterization data and the first target domain feature characterization data; and adjusting parameters of the cross-domain feature characterization model based on the difference value. In adjusting the parameters of the cross-domain feature characterization model, a back propagation algorithm or a random gradient descent algorithm may be used to adjust the parameters of the cross-domain feature characterization model. The trained first target domain feature characterization model may be a neural network model formed by a plurality of fully connected layers, for example, a neural network model formed by three fully connected layers, a first fully connected layer may map input data to 1028-dimensional vectors, a second fully connected layer may map 1028-dimensional vectors to 512-dimensional vectors, and a third fully connected layer may map 512-dimensional vectors to 128-dimensional vectors. The target loss function can be any loss function such as a cross entropy loss function, a softmax loss function, an L1 loss function, and an L2 loss function. The second cross-domain feature characterization data may be a cross-domain feature characterization vector, and the first target domain feature characterization data may be a target domain feature characterization vector. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In a specific example, by determining a difference value between the second cross-domain feature characterization data and the first target-domain feature characterization data, the currently obtained second cross-domain feature characterization data is evaluated to be used as a basis for subsequently training the cross-domain feature characterization model. Specifically, the difference values may be transmitted back to the cross-domain feature characterization model, thereby iteratively training the cross-domain feature characterization model. The training of the cross-domain feature characterization model is an iterative process, and the embodiment of the present application only describes one training process, but it should be understood by those skilled in the art that this training mode may be adopted for each training of the cross-domain feature characterization model until the training of the cross-domain feature characterization model is completed. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In some optional embodiments, when the cross-domain feature characterization model to be trained is trained, the cross-domain feature characterization model to be trained is trained based on the first feature data, the second feature data, fourth feature data of a sample object associated with the sample user in the target domain, and behavior labeling data of the sample user for the sample object. Thereby, the trained cross-domain feature characterization model is enabled to generate cross-domain feature characterization data of the target domain feature-missing user about the object in the target domain. It can be understood that any embodiment for training the cross-domain feature characterization model to be trained is applicable to this embodiment, and this is not limited in this embodiment of the present application.
In the embodiment of the present application, the sample object may be understood as an object, such as a commodity, a literary work, a song, music, a movie, etc., which the sample user relates to in the user log data in the target field. The fourth feature data may be understood as overall feature data of the sample object in the target field, for example, category data, seller information, keyword information, sales information, and the like of the sample object may be included. The behavior labeling data can comprise labeling data for representing that a sample user browses and clicks a sample object, labeling data for representing that the sample user browses but does not click the sample object, labeling data for representing that the sample user clicks and purchases the sample object, labeling data for representing that the sample user clicks but does not purchase the sample object, and the like. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In some optional embodiments, when the cross-domain feature characterization model to be trained is trained based on the first feature data, the second feature data, fourth feature data of a sample object associated with the sample user in the target domain, and behavior labeling data of the sample user for the sample object, first training is performed on the cross-domain feature characterization model and a first target domain feature characterization model based on the first feature data and the second feature data; performing second training on the first target field characteristic representation model, the first object characteristic representation model and the first object estimation model based on the second characteristic data, the fourth characteristic data and the behavior marking data; and performing third training on the cross-domain feature characterization model, the first object feature characterization model and the first object pre-estimation model based on the first feature data, the fourth feature data and the behavior labeling data. Therefore, the trained cross-domain feature characterization model can generate cross-domain feature characterization data of the target domain object estimation task of the user with the missing target domain features. It should be understood that any embodiment that trains the cross-domain feature characterization model to be trained based on the first feature data, the second feature data, the fourth feature data of the sample object associated with the sample user in the target domain, and the behavior labeling data of the sample user for the sample object may be applied to this embodiment, which is not limited in this embodiment of the present application.
In this embodiment, the first object feature characterization model may be a neural network model composed of a plurality of fully-connected layers, for example, a neural network model composed of three fully-connected layers, a first fully-connected layer may map input data to 1028-dimensional vectors, a second fully-connected layer may map 1028-dimensional vectors to 512-dimensional vectors, and a third fully-connected layer may map 512-dimensional vectors to 128-dimensional vectors. The first object prediction model may be a neural network model composed of a plurality of fully connected layers and a computation layer, for example, a neural network model composed of three fully connected layers and a computation layer, the first fully connected layer may map input data into 64-dimensional vectors, the second fully connected layer may map 64-dimensional vectors into 32-dimensional vectors, the third fully connected layer may map 32-dimensional vectors into 1-dimensional vectors, and the computation layer may normalize the 1-dimensional vectors. In addition, when the cross-domain feature characterization model to be trained is trained, the output end of the cross-domain feature characterization model and the output end of the first target domain feature characterization model are respectively connected with the input end of the calculation layer for measuring the distance, the output end of the cross-domain feature characterization model and the output end of the first target feature characterization model are respectively connected with the input end of the first target pre-estimation model, and the output end of the first target domain feature characterization model and the output end of the first target feature characterization model are respectively connected with the input end of the first target pre-estimation model. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In some optional embodiments, during the first training of the cross-domain feature characterization model and the first target domain feature characterization model, generating, by the cross-domain feature characterization model, second cross-domain feature characterization data of the sample user in the target domain based on the first feature data; generating first target domain feature characterization data of the sample user based on the second feature data through the first target domain feature characterization model; and performing first training on the cross-domain feature characterization model and the first target domain feature characterization model based on the second cross-domain feature characterization data and the first target domain feature characterization data. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In a specific example, when second cross-domain feature characterization data of the sample user in the target domain is generated, mapping operation is performed on the first feature data through the cross-domain feature characterization model to generate the second cross-domain feature characterization data of the sample user in the target domain. When the first target domain feature characterization data of the sample user is generated, mapping operation is carried out on the second feature data through the first target domain feature characterization model, so that the first target domain feature characterization data of the sample user is generated. Determining distance values of the second cross-domain feature characterization data and the first target domain feature characterization data through a distance metric function when the first training is performed on the cross-domain feature characterization model and the first target domain feature characterization model based on the second cross-domain feature characterization data and the first target domain feature characterization data; adjusting parameters of the cross-domain feature characterization model and the first target domain feature characterization model based on the distance value. Determining a difference value between the second cross-domain feature characterization data and the first target domain feature characterization data based on the distance value through a target loss function when adjusting parameters of the cross-domain feature characterization model and the first target domain feature characterization model; adjusting parameters of the cross-domain feature characterization model and the first target domain feature characterization model based on the difference value. In particular, a back propagation algorithm, or a random gradient descent algorithm, may be employed to adjust the parameters of the cross-domain feature characterization model and the first target domain feature characterization model. The distance metric function includes, but is not limited to, euclidean distance, a distance metric learning function, a discriminator in a generative countermeasure network, and the like. The target loss function can be any loss function such as a cross entropy loss function, a softmax loss function, an L1 loss function, and an L2 loss function. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In a specific example, by determining loss values of the second cross-domain feature characterization data and the first target domain feature characterization data, the currently obtained second cross-domain feature characterization data is evaluated as a basis for subsequently training the cross-domain feature characterization model and the first target domain feature characterization model. Specifically, the loss value may be transmitted back to the cross-domain feature characterization model and the first target domain feature characterization model, thereby iteratively training the cross-domain feature characterization model and the first target domain feature characterization model. The training of the cross-domain feature characterization model and the first target-domain feature characterization model is an iterative process, and the embodiment of the present application only describes one training process, but it should be understood by those skilled in the art that this training mode may be adopted for each training of the cross-domain feature characterization model and the first target-domain feature characterization model until the training of the cross-domain feature characterization model and the first target-domain feature characterization model is completed. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In some optional embodiments, when performing the second training on the first target domain feature characterization model, the first object feature characterization model, and the first object pre-estimation model, generating, by the first target domain feature characterization model, first target domain feature characterization data of the sample user based on the second feature data; generating, by the first object characterization model, characterization data for the sample object based on the fourth characterization data; estimating the behavior of the sample user aiming at the sample object through the first object estimation model based on the first target field characteristic data and the characteristic data of the sample object so as to obtain behavior estimation data of the sample user aiming at the sample object; and performing second training on the first target field characteristic representation model, the first object characteristic representation model and the first object estimation model based on the behavior marking data and the behavior estimation data. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In a specific example, when the first target domain feature characterization data of the sample user is generated, the second feature data is subjected to a mapping operation to generate the first target domain feature characterization data of the sample user. When generating the feature characterization data of the sample object, performing a mapping operation on the fourth feature data to generate the feature characterization data of the sample object. Wherein the feature characterization data of the sample object may be a feature characterization vector of the sample object. When behavior prediction data of the sample user for the sample object is obtained, mapping operation and normalization operation are carried out on the first target field characteristic data and the characteristic characterization data of the sample object, so that the behavior prediction data of the sample user for the sample object is obtained. The behavior prediction data can comprise prediction data representing that a sample user browses and clicks a sample object, prediction data representing that the sample user browses and does not click the sample object, prediction data representing that the sample user clicks and purchases the sample object, prediction data representing that the sample user clicks and does not purchase the sample object, and the like. When the first target field characteristic representation model, the first object characteristic representation model and the first object estimation model are subjected to second training based on the behavior labeling data and the behavior estimation data, determining a difference value between the behavior labeling data and the behavior estimation data through a target loss function; and adjusting parameters of the first target field characteristic model, the first object characteristic model and the first object pre-estimation model based on the difference value. In adjusting the parameters of the first target domain characteristic model, the first object characteristic model and the first object pre-estimation model, a back propagation algorithm or a random gradient descent algorithm may be adopted to adjust the parameters of the first target domain characteristic model, the first object characteristic model and the first object pre-estimation model. The target loss function can be any loss function such as a cross entropy loss function, a softmax loss function, an L1 loss function, and an L2 loss function. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In a specific example, the behavior estimation data obtained currently is evaluated by determining a difference value between the behavior labeling data and the behavior estimation data, so as to serve as a basis for subsequently training the first target domain feature characterization model, the first object feature characterization model and the first object estimation model. Specifically, the difference values may be transmitted to the first target domain feature representation model, the first object feature representation model, and the first object prediction model in a reverse direction, so as to iteratively train the first target domain feature representation model, the first object feature representation model, and the first object prediction model. The training of the first target domain feature characterization model, the first object feature characterization model and the first object estimation model is an iterative process, and the embodiment of the present application only describes one training process, but it should be understood by those skilled in the art that this training mode may be adopted for each training of the first target domain feature characterization model, the first object feature characterization model and the first object estimation model until the training of the first target domain feature characterization model, the first object feature characterization model and the first object estimation model is completed. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In some optional embodiments, in the third training of the cross-domain feature characterization model, the first object feature characterization model and the first object pre-estimation model, second cross-domain feature characterization data of the sample user in the target domain is generated through the cross-domain feature characterization model based on the first feature data; generating, by the first object characterization model, characterization data for the sample object based on the fourth characterization data; estimating the behavior of the sample user aiming at the sample object by the first object estimation model based on the second cross-domain feature characterization data and the feature characterization data of the sample object to obtain behavior estimation data of the sample user aiming at the sample object; and performing third training on the cross-domain feature characterization model, the first object feature characterization model and the first object prediction model based on the behavior labeling data and the behavior prediction data. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In a specific example, when second cross-domain feature characterization data of the sample user in the target domain is generated, mapping operation is performed on the first feature data through the cross-domain feature characterization model to generate the second cross-domain feature characterization data of the sample user in the target domain. When generating the feature characterization data of the sample object, performing a mapping operation on the fourth feature data to generate the feature characterization data of the sample object. When behavior prediction data of the sample user for the sample object is obtained, mapping operation and normalization operation are carried out on the second cross-domain feature characterization data and the feature characterization data of the sample object, so that the behavior prediction data of the sample user for the sample object is obtained. When the cross-domain feature characterization model, the first object feature characterization model and the first object prediction model are subjected to third training based on the behavior labeling data and the behavior prediction data, determining a difference value between the behavior labeling data and the behavior prediction data through a target loss function; and adjusting parameters of the cross-domain feature characterization model, the first object feature characterization model and the first object pre-estimation model based on the difference values. In adjusting the parameters of the cross-domain feature characterization model, the first object feature characterization model and the first object pre-estimation model, a back propagation algorithm or a random gradient descent algorithm may be adopted to adjust the parameters of the cross-domain feature characterization model, the first object feature characterization model and the first object pre-estimation model. The target loss function can be any loss function such as a cross entropy loss function, a softmax loss function, an L1 loss function, and an L2 loss function. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In a specific example, the behavior estimation data obtained currently is evaluated by determining a difference value between the behavior labeling data and the behavior estimation data, so as to serve as a basis for subsequently training the cross-domain feature characterization model, the first object feature characterization model and the first object estimation model. Specifically, the difference values may be transmitted to the cross-domain feature characterization model, the first object feature characterization model, and the first object pre-estimation model in a reverse direction, so as to iteratively train the cross-domain feature characterization model, the first object feature characterization model, and the first object pre-estimation model. The training of the cross-domain feature characterization model, the first object feature characterization model and the first object estimation model is an iterative process, and the embodiment of the present application only describes one training process, but it should be understood by those skilled in the art that this training mode may be adopted for each training of the cross-domain feature characterization model, the first object feature characterization model and the first object estimation model until the training of the cross-domain feature characterization model, the first object feature characterization model and the first object estimation model is completed. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In a specific example, u is shown in FIG. 2sFirst feature data, u, representing active users in the source domaintRepresenting second feature data of the active users in the target field, i representing fourth feature data of sample objects associated with the sample users in the target field, G representing a cross-field feature characterization model, fuRepresenting a first target domain feature characterization model, fiRepresenting a first object feature characterization model, S representing a computation layer for measuring distance, h representing a first object prediction model, LsRepresenting a first target loss function, LtRepresenting a second target loss function, LcRepresenting a third objective loss function. Second cross-domain feature characterization data output by cross-domain feature characterization model G and first target domain feature characterization model fuInputting the output first target domain feature characterization data into a calculation layer S to calculate a distance value between the second cross-domain feature characterization data and the first target domain feature characterization data, and then obtaining a first target loss function LsBased on the distance values, respective disparity values are determined. Second cross-domain feature characterization data output by the cross-domain feature characterization model G and first object feature characterization model fiInputting the output characteristic characterization data of the sample object into a first object prediction model h to obtain behavior prediction data of a sample user for the sample object, and then performing a third target loss function LcAnd calculating the difference value of the behavior marking data and the behavior estimation data. First target domain feature characterization model fuThe output first target domain feature characterization data and a first object feature characterization model fiInputting the output characteristic characterization data of the sample object into a first object prediction model h to obtain behavior prediction data of a sample user for the sample object, and then performing a second target loss function LtAnd calculating the difference value of the behavior marking data and the behavior estimation data. It is to be understood that the above description is intended to be illustrative only,the embodiment of the present application is not limited to this.
As can be seen from the above description, the training process of the cross-domain feature characterization model to be trained includes a first training, a second training, and a third training. Accordingly, the first trained objective loss function may be referred to as a first objective loss function, the second trained objective loss function may be referred to as a second objective loss function, and the third trained objective loss function may be referred to as a third objective loss function. Wherein the first target loss function represents a loss value of a similarity measure between the cross-domain feature representation and the target-domain feature representation, the effect of which is to make the cross-domain feature representation more similar to the real target-domain feature representation. The second target loss function represents a loss value for predicting the behavior of the sample user for the sample object under the target field characteristic representation, and the second target loss function has the function of enabling the real target field characteristic representation, namely the target object learned by the cross-field characteristic representation model, to have the capability of well predicting the behavior of the user for the object. The third target loss function represents a loss value for predicting the behavior of the sample user for the sample object under the cross-domain feature characterization, and the third target loss function has the function of restraining the cross-domain feature characterization to have the capability of predicting the behavior of the user for the object. In a specific embodiment, the first target loss function, the second target loss function, and the third target loss function may be weighted and summed to obtain an overall target loss function for training the cross-domain feature characterization model. The first, second, and third training of the cross-domain feature characterization model may then be controlled by an optimizer based on the overall objective loss function. In addition, the first object prediction model in the second training and the first object prediction model in the third training may be the same object prediction model, or of course, may be different object prediction models. When the degree of approximation of the distribution of the source field and the distribution of the target field is high enough, the learning of the cross-field feature representation is similar to the learning of the target field feature representation, the first object estimation model in the second training and the first object estimation model in the third training can be the same object estimation model, namely the parameters of the first object estimation model are shared, and therefore the training speed of the cross-field feature representation model can be increased. For example, when the source domain is a tianmao client and the target domain is a panning client, the degree of approximation of the distribution of the source domain and the distribution of the target domain is sufficiently high, and the first object prediction model in the second training and the first object prediction model in the third training may be the same object prediction model. When the difference between the distribution of the source field and the distribution of the target field is large, the learning of the cross-field feature characterization is not similar to the learning of the target field feature characterization, the first object estimation model in the second training and the first object estimation model in the third training may be different object estimation models, that is, the first object estimation model in the second training and the first object estimation model in the third training are two independent object estimation models. For example, when the source domain is a high-end client and the target domain is a panning client, the difference between the distribution of the source domain and the distribution of the target domain is large enough, and the first object prediction model in the second training and the first object prediction model in the third training may be different object prediction models. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In step S102, based on the third feature data of the user to be characterized in the source domain, generating first cross-domain feature characterization data of the user to be characterized in the target domain through the trained cross-domain feature characterization model.
In the embodiment of the application, the trained cross-domain feature characterization model has the function of migrating the knowledge of the active users to the users to be characterized so as to make up for the target domain features which are insufficient for the users to be characterized. That is, when there are sample users with source domain feature data similar to the third feature data of the current user to be characterized in the source domain among the sample users (active users) in step S101, the feature characterization of these sample users in the target domain will be used to generate the cross-domain feature characterization of the user to be characterized. The users to be characterized include users with different degrees of missing features of the target domain, for example, cold-start users, which can be understood as users with features that are severely missing in the target domain and abundant in the source domain. The third feature data can be understood as overall feature data of the user to be characterized in the source field, including behavior feature data and user portrait data of the user to be characterized in the source field. The first cross-domain feature characterization data can be understood as data obtained by migrating third feature data of a user to be characterized in the source domain and characterizing the migrated feature data by using a feature characterization mode of a target domain obtained by training. Therefore, the understanding of the user to be characterized in the target field can be deepened, and the personalized characterization effect of the user to be characterized in the target field is further improved.
In a specific example, when first cross-domain feature characterization data of the user to be characterized in the target domain is generated, mapping operation is performed on third feature data of the user to be characterized in the source domain through the trained cross-domain feature characterization model, so as to generate the first cross-domain feature characterization data of the user to be characterized in the target domain. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In some optional embodiments, when the first cross-domain feature characterization data of the user to be characterized in the target domain is generated based on the third feature data of the user to be characterized in the source domain through the trained cross-domain feature characterization model, the third cross-domain feature characterization data of the user to be characterized for the object pre-estimation task in the target domain is generated based on the third feature data of the user to be characterized in the source domain through the trained cross-domain feature characterization model. The object prediction tasks comprise object click rate prediction tasks or object conversion rate prediction tasks and the like, and the third cross-domain feature characterization data can be cross-domain feature characterization vectors. Therefore, the personalized characterization effect of the user to be characterized on the object estimation task in the target field can be effectively improved. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In a specific example, when third cross-domain feature characterization data of the to-be-characterized user for the object pre-estimation task in the target domain is generated, mapping operation is performed on the third feature data of the to-be-characterized user in the source domain through the trained cross-domain feature characterization model, so as to generate the third cross-domain feature characterization data of the to-be-characterized user for the object pre-estimation task in the target domain. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In practical applications, in a recommendation or search scenario (such as recommendation or search of the kyoto and amazon) of a cold start user of a product such as an electronic commerce and the like, there may be a place where a cross-domain feature characterization learning mechanism is utilized to improve a user characterization effect. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
According to the user characterization method provided by the embodiment of the application, the cross-domain feature characterization model to be trained is trained based on the first feature data of the sample user in the source domain and the second feature data of the sample user in the target domain, the first cross-domain feature characterization data of the user to be characterized in the target domain is generated based on the third feature data of the user to be characterized in the source domain through the trained cross-domain feature characterization model, compared with the existing other modes, the cross-domain feature characterization data of the user to be characterized in the target domain is generated based on the feature data of the user to be characterized in the source domain, and the personalized characterization effect of the user with different degrees of missing features in the target domain can be effectively improved.
The user characterization method of the present embodiment may be performed by any suitable device having data processing capabilities, including but not limited to: cameras, terminals, mobile terminals, PCs, servers, in-vehicle devices, entertainment devices, advertising devices, Personal Digital Assistants (PDAs), tablet computers, notebook computers, handheld game consoles, smart glasses, smart watches, wearable devices, virtual display devices or display enhancement devices (such as Google Glass, Oculus rise, Hololens, Gear VR), and the like.
Referring to fig. 3, a flowchart illustrating steps of a user characterization method according to the second embodiment of the present application is shown.
Specifically, the user characterization method provided by this embodiment includes the following steps:
in step S201, the cross-domain feature characterization model to be trained is trained based on the first feature data, the second feature data, fourth feature data of a sample object associated with the sample user in the target domain, and behavior labeling data of the sample user for the sample object.
Since the specific implementation manner of step S201 is similar to the specific implementation manner of training the cross-domain feature characterization model to be trained in the first embodiment, details are not repeated here.
In step S202, based on the third feature data of the to-be-characterized user in the source domain, generating third cross-domain feature characterization data of the to-be-characterized user for the object pre-estimation task in the target domain through the trained cross-domain feature characterization model.
Since the specific implementation manner of step S202 is similar to the specific implementation manner of generating the third cross-domain feature characterization data of the object pre-estimation task for the target domain by the user to be characterized in the first embodiment, details are not repeated here.
In step S203, source domain feature characterization data of the to-be-characterized user for the object pre-estimation task is generated based on third feature data of the to-be-characterized user in the source domain through a source domain feature characterization model.
In the embodiment of the present application, the source domain feature characterization model may be a neural network model composed of a plurality of fully-connected layers, for example, a neural network model composed of three fully-connected layers, a first fully-connected layer may map input data to 1028-dimensional vectors, a second fully-connected layer may map 1028-dimensional vectors to 512-dimensional vectors, and a third fully-connected layer may map 512-dimensional vectors to 128-dimensional vectors. The source domain feature characterization data is specifically a source domain feature characterization vector. It should be noted that, although the data input by the cross-domain feature characterization model and the source-domain feature characterization model are the third feature data of the user to be characterized in the source domain, their roles are different. Specifically, the source domain feature characterization model focuses on the third feature data, and the cross-domain feature characterization model mainly generates third cross-domain feature characterization data of the to-be-characterized user for the object estimation task in the target domain by using the third feature data of the to-be-characterized user in the source domain. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In some optional embodiments, when source domain feature characterization data of the to-be-characterized user for the object pre-estimation task is generated, a mapping operation is performed on third feature data of the to-be-characterized user in the source domain through a source domain feature characterization model, so as to generate source domain feature characterization data of the to-be-characterized user for the object pre-estimation task. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In step S204, second target domain feature characterization data of the to-be-characterized user for the object pre-estimation task is generated based on fifth feature data of the to-be-characterized user in the target domain through a second target domain feature characterization model.
In this embodiment, the second target domain feature characterization model may be a neural network model composed of a plurality of fully-connected layers, for example, a neural network model composed of three fully-connected layers, a first fully-connected layer may map input data to 1028-dimensional vectors, a second fully-connected layer may map 1028-dimensional vectors to 512-dimensional vectors, and a third fully-connected layer may map 512-dimensional vectors to 128-dimensional vectors. The fifth feature data can be understood as the overall feature data of the user to be characterized in the target field, including the behavior feature data and the user portrait data of the user to be characterized in the target field. The second target domain feature characterization data may be a target domain feature characterization vector. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In some optional embodiments, when second target domain feature characterization data of the to-be-characterized user for the object pre-estimation task is generated, a mapping operation is performed on fifth feature data of the to-be-characterized user in the target domain through a second target domain feature characterization model, so as to generate second target domain feature characterization data of the to-be-characterized user for the object pre-estimation task. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In step S205, based on the third cross-domain feature characterization data, the source domain feature characterization data, and the second target domain feature characterization data, comprehensive feature characterization data of the user to be characterized is determined.
In an embodiment of the present application, the integrated feature characterization data may be an integrated feature characterization vector. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In some optional embodiments, when determining the comprehensive feature characterization data of the user to be characterized, processing the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data through a domain attention model to obtain weight values corresponding to the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data respectively; determining comprehensive feature characterization data of the user to be characterized based on weighted values respectively corresponding to the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data, the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data. It is understood that any embodiment of determining the comprehensive feature characterization data of the user to be characterized may be applied to this embodiment, and this is not limited in this embodiment of the present application.
In a specific embodiment, for different users to be characterized, the target domain feature of some users to be characterized is very seriously lost, the target domain feature of some users to be characterized is not very serious, and the requirements of the users to be characterized on cross-domain feature are different. Therefore, the attention degree of the user to be characterized on the source field characteristic characterization, the target field characteristic characterization and the cross-field characteristic characterization and the attention degree of the three user characteristic characterizations are calculated through the domain attention model, and then the final characteristic characterization of the user to be characterized is obtained through weighting. For a user to be characterized with high activity, the attention of the target domain feature characterization is high, and the attention of the source domain feature characterization and the attention of the cross-domain feature characterization are low. For users with some missing target domain features, the attention of the source domain feature characterization and the cross-domain feature characterization is high, and the importance degree of the cross-domain feature characterization is higher than that of the source domain feature characterization. For the user to be characterized with completely missing target domain features, the importance degree of the cross-domain feature characterization is far higher than that of the source domain feature characterization. Therefore, the cross-domain feature characterization, the source domain feature characterization and the target domain feature characterization are focused to different degrees through the domain attention model, and the features of the user to be characterized can be more accurately characterized, so that the personalized characterization effect of the user to be characterized in the target domain is further improved. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In one particular example, the domain attention model includes a convolutional layer, a fully-connected layer connected to an output of the convolutional layer, and a computational layer connected to an output of the fully-connected layer. When the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data are processed through a domain attention model to obtain weight values corresponding to the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data respectively, performing convolution operation on the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data through a convolution layer of the domain attention model to obtain feature maps of the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data; performing mapping operation on the feature map through a full connection layer of the domain attention model to obtain weight feature vectors of the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data; and performing normalization operation on the weight feature vector through a calculation layer of the domain attention model to obtain weight values corresponding to the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data respectively. Therefore, the weight values corresponding to the third cross-domain feature representation data, the source domain feature representation data and the second target domain feature representation data can be accurately obtained. It can be understood that any embodiment of obtaining weight values corresponding to the third cross-domain feature characterization data, the source domain feature characterization data, and the second target domain feature characterization data may be applied to this embodiment, and this is not limited in this embodiment of the present application.
In a specific example, when determining the comprehensive characteristic representation data of the user to be characterized based on the weight values respectively corresponding to the third cross-domain characteristic representation data, the source domain characteristic representation data and the second target domain characteristic representation data, the third cross-domain characteristic representation data and the weight values corresponding to the third cross-domain characteristic representation data are multiplied to obtain a first multiplication result; multiplying the source field characteristic representation data by a weight value corresponding to the source field characteristic representation data to obtain a second multiplication result; multiplying the second target field characteristic data by a weight value corresponding to the second target field characteristic data to obtain a third multiplication result; and adding the first multiplication result, the second multiplication result and the third multiplication result to obtain comprehensive characteristic characterization data of the user to be characterized. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In some optional embodiments, the method further comprises: generating characteristic characterization data of the object to be pre-estimated aiming at the object pre-estimation task based on sixth characteristic data of the object to be pre-estimated in the target field through a second object characteristic characterization model; estimating the behavior of the user to be characterized aiming at the object to be estimated through a second object estimation model based on the comprehensive characteristic characterization data of the user to be characterized and the characteristic characterization data of the object to be estimated aiming at the object estimation task, so as to obtain the behavior estimation data of the user to be characterized aiming at the object to be estimated. Therefore, the user to be characterized is characterized by the comprehensive characteristic characterization data, and the behavior of the user to be characterized aiming at the object to be estimated can be accurately estimated. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In a specific example, the second object feature characterization model may be a neural network model composed of a plurality of fully connected layers, for example, a neural network model composed of three fully connected layers, a first fully connected layer may map input data to 1028-dimensional vectors, a second fully connected layer may map 1028-dimensional vectors to 512-dimensional vectors, and a third fully connected layer may map 512-dimensional vectors to 128-dimensional vectors. The sixth feature data may be understood as the overall feature data of the object to be estimated in the target field, for example, category data, seller information, keyword information, sales information, and the like of the object to be estimated may be included. The objects to be estimated comprise commodities, literary works, TV plays, music, movies and the like in the target field. The second object prediction model may be a neural network model composed of a plurality of fully connected layers and a computation layer, for example, a neural network model composed of three fully connected layers and a computation layer, a first fully connected layer may map input data into 64-dimensional vectors, a second fully connected layer may map 64-dimensional vectors into 32-dimensional vectors, a third fully connected layer may map 32-dimensional vectors into 1-dimensional vectors, and a computation layer may normalize the 1-dimensional vectors. The behavior of the user to be represented aiming at the object to be estimated comprises browsing and clicking the object to be estimated by the user to be represented, browsing and not clicking the object to be estimated by the user to be represented, clicking and purchasing the object to be estimated by the user to be represented, clicking and not purchasing the object to be estimated by the user to be represented and the like. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In a specific example, u is shown in FIG. 4sRepresenting third characteristic data, u, of the user to be characterized in the source domaintAnd the fifth characteristic data represent the fifth characteristic data of the user to be characterized in the target field, and the i represents the sixth characteristic data of the object to be estimated in the target field. The output end of the trained cross-domain feature characterization model is respectively connected with the input end of the domain attention model and the input end of the calculation layer for calculating the weighted sum, the output end of the source domain feature characterization model is respectively connected with the input end of the domain attention model and the input end of the calculation layer for calculating the weighted sum, the output end of the second target domain feature characterization model is respectively connected with the input end of the domain attention model and the input end of the calculation layer for calculating the weighted sum, and the output end of the calculation layer for calculating the weighted sum and the output end of the second target feature characterization model are respectively connected with the input end of the second target pre-estimation model. Inputting third cross-domain feature characterization data output by the trained cross-domain feature characterization model, source domain feature characterization data output by the source domain feature characterization model and second target domain feature characterization data output by the second target domain feature characterization modelAnd entering a domain attention model to obtain weight values corresponding to the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data respectively. Then, respectively inputting the weight values, the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data corresponding to the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data into a calculation layer for calculating a weighted sum so as to obtain comprehensive feature characterization data of the user to be characterized. And finally, inputting the comprehensive characteristic representation data of the user to be represented and the characteristic representation data of the object to be estimated, which is output by the second object characteristic representation model, aiming at the object estimation task into the second object estimation model so as to obtain behavior estimation data of the user to be represented aiming at the object to be estimated. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In a specific example, the framework shown in FIG. 4 needs to be trained before the framework shown in FIG. 4 processes the object pre-estimation task. Specifically, in the training process, parameters of the trained cross-domain feature characterization model are ensured to be fixed, and parameters of the source domain feature characterization model, the second target domain feature characterization model, the domain attention model, the second object feature characterization model and the second object estimation model are adjusted. More specifically, the source field feature characterization model to be trained, the second target field feature characterization model, the domain attention model, the second object feature characterization model and the second object pre-estimation model are trained based on feature data of any user in a source field, feature data of any user in a target field, feature data of a sample object associated with any user in the target field, and behavior labeling data of any user for the sample object. More specifically, generating cross-domain feature characterization data of an object estimation task of any user in a target domain based on feature data of any user in a source domain through a trained cross-domain feature characterization model; generating source field characteristic representation data of an object estimation task of any user in a target field based on the characteristic data of any user in the source field through a source field characteristic representation model to be trained; generating target field characteristic data of an object estimation task of any user in the target field based on the characteristic data of any user in the target field through a second target field characteristic model to be trained; processing the cross-domain feature characterization data, the source domain feature characterization data and the target domain feature characterization data through a domain attention model to be trained to obtain weighted values corresponding to the cross-domain feature characterization data, the source domain feature characterization data and the target domain feature characterization data respectively; determining comprehensive feature characterization data of any user based on the cross-domain feature characterization data, the source domain feature characterization data and the target domain feature characterization data, and weight values respectively corresponding to the cross-domain feature characterization data, the source domain feature characterization data and the target domain feature characterization data; generating feature characterization data of the sample object for an object pre-estimation task in the target field based on feature data of the sample object associated with the arbitrary user in the target field through a second object feature characterization model to be trained; estimating the behavior of any user aiming at the sample object through a second object estimation model to be trained based on the comprehensive characteristic representation data of any user and the characteristic representation data of the object estimation task of the sample object aiming at the target field, so as to obtain the behavior estimation data of any user aiming at the sample object; and training the source field characteristic characterization model to be trained, the second target field characteristic characterization model, the domain attention model, the second object characteristic characterization model and the second object estimation model based on the behavior marking data and the behavior estimation data of the sample object by the arbitrary user. More specifically, a difference value between behavior marking data of the arbitrary user for the sample object and the behavior prediction data is determined through a target loss function; and adjusting parameters of the source field characteristic model to be trained, the second target field characteristic model, the domain attention model, the second object characteristic model and the second object pre-estimation model based on the difference value. The target loss function can be any loss function such as a cross entropy loss function, a softmax loss function, an L1 loss function, and an L2 loss function. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
In a specific example, the behavior estimation data obtained currently is evaluated by determining a difference value between the behavior labeling data and the behavior estimation data, so as to serve as a basis for subsequently training the source domain feature characterization model, the second target domain feature characterization model, the domain attention model, the second object feature characterization model and the second object estimation model. Specifically, the difference value may be reversely transmitted to the source domain feature characterization model to be trained, the second target domain feature characterization model, the domain attention model, the second object feature characterization model, and the second object prediction model, so as to iteratively train the source domain feature characterization model to be trained, the second target domain feature characterization model, the domain attention model, the second object feature characterization model, and the second object prediction model. The training of the source domain feature characterization model, the second target domain feature characterization model, the domain attention model, the second object feature characterization model and the second object prediction model is an iterative process, and the embodiment of the present application only describes one training process, but it should be understood by those skilled in the art that this training mode may be adopted for each training of the source domain feature characterization model, the second target domain feature characterization model, the domain attention model, the second object feature characterization model and the second object prediction model until the training of the source domain feature characterization model, the second target domain feature characterization model, the domain attention model, the second object feature characterization model and the second object prediction model is completed. It should be understood that the above description is only exemplary, and the embodiments of the present application are not limited in this respect.
According to the user characterization method provided by the embodiment of the application, based on first feature data of a sample user in a source field, second feature data of the sample user in a target field, fourth feature data of a sample object associated with the sample user in the target field and behavior marking data of the sample user for the sample object, a cross-domain feature characterization model to be trained is trained, and through the trained cross-domain feature characterization model, based on third feature data of the user to be characterized in the source field, third cross-domain feature characterization data of the user to be characterized for an object estimation task in the target field is generated The individuation of affairs characterizes the effect.
In addition, source domain feature characterization data of the user to be characterized aiming at the object estimation task in the target domain is generated through a source domain feature characterization model based on third feature data of the user to be characterized in the source domain, second target domain feature characterization data of the user to be characterized aiming at the object estimation task in the target domain is generated through a second target domain feature characterization model based on fifth feature data of the user to be characterized in the target domain, comprehensive feature characterization data of the user to be characterized is determined based on the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data, and compared with the existing other modes, the comprehensive feature characterization data of the user to be characterized is determined based on the cross-domain feature characterization data, the source domain feature characterization data and the target domain feature characterization data, the personalized characterization effect of the user with different missing characteristics of the target field on the object estimation task in the target field can be further improved.
The user characterization method of the present embodiment may be performed by any suitable device having data processing capabilities, including but not limited to: cameras, terminals, mobile terminals, PCs, servers, in-vehicle devices, entertainment devices, advertising devices, Personal Digital Assistants (PDAs), tablet computers, notebook computers, handheld game consoles, smart glasses, smart watches, wearable devices, virtual display devices or display enhancement devices (such as Google Glass, Oculus rise, Hololens, Gear VR), and the like.
Referring to fig. 5, a schematic structural diagram of a user characterization device in the third embodiment of the present application is shown.
The user characterization device of the embodiment comprises: the first training module 301 is configured to train a to-be-trained cross-domain feature characterization model based on first feature data of a sample user in a source domain and second feature data of the sample user in a target domain; a first generating module 302, configured to generate, through the trained cross-domain feature characterization model, first cross-domain feature characterization data of the user to be characterized in the target domain based on third feature data of the user to be characterized in the source domain.
The user characterization device of this embodiment is used to implement the corresponding user characterization method in the foregoing multiple method embodiments, and has the beneficial effects of the corresponding method embodiments, which are not described herein again.
Referring to fig. 6, a schematic structural diagram of a user characterization device in the fourth embodiment of the present application is shown.
The user characterization device of the embodiment comprises: the first training module 401 is configured to train a cross-domain feature characterization model to be trained based on first feature data of a sample user in a source domain and second feature data of the sample user in a target domain; a first generating module 402, configured to generate, through the trained cross-domain feature characterization model, first cross-domain feature characterization data of the user to be characterized in the target domain based on third feature data of the user to be characterized in the source domain.
Optionally, the first training module 401 includes: a second training module 4011, configured to train the cross-domain feature characterization model to be trained based on the first feature data, the second feature data, fourth feature data of a sample object associated with the sample user in the target domain, and behavior labeling data of the sample user for the sample object.
Optionally, the second training module 4011 includes: a third training module 4012, configured to perform first training on the cross-domain feature characterization model and the first target domain feature characterization model based on the first feature data and the second feature data; a fourth training module 4016, configured to perform second training on the first target domain feature characterization model, the first object feature characterization model, and the first object prediction model based on the second feature data, the fourth feature data, and the behavior labeling data; a fifth training module 4021, configured to perform third training on the cross-domain feature characterization model, the first object feature characterization model, and the first object prediction model based on the first feature data, the fourth feature data, and the behavior labeling data.
Optionally, the third training module 4012 includes: a second generating module 4013, configured to generate, by the cross-domain feature characterization model, second cross-domain feature characterization data of the sample user in the target domain based on the first feature data; a third generating module 4014, configured to generate, by the first target domain feature characterization model, first target domain feature characterization data of the sample user based on the second feature data; a sixth training module 4015, configured to perform first training on the cross-domain feature characterization model and the first target domain feature characterization model based on the second cross-domain feature characterization data and the first target domain feature characterization data.
Optionally, the sixth training module 4015 is specifically configured to: determining a distance value of the second cross-domain feature characterization data and the first target domain feature characterization data by a distance metric function; adjusting parameters of the cross-domain feature characterization model and the first target domain feature characterization model based on the distance value.
Optionally, the fourth training module 4016 includes: a fourth generating module 4017, configured to generate, by using the first target domain feature characterization model, first target domain feature characterization data of the sample user based on the second feature data; a fifth generating module 4018, configured to generate, by the first object feature characterization model, feature characterization data of the sample object based on the fourth feature data; the first estimation module 4019 is configured to estimate, by using the first object estimation model, a behavior of the sample user with respect to the sample object based on the first target domain feature characterization data and the feature characterization data of the sample object, so as to obtain behavior estimation data of the sample user with respect to the sample object; a seventh training module 4020, configured to perform second training on the first target domain feature characterization model, the first object feature characterization model, and the first object prediction model based on the behavior labeling data and the behavior prediction data.
Optionally, the seventh training module 4020 is specifically configured to: determining a difference value between the behavior marking data and the behavior estimation data through a target loss function; and adjusting parameters of the first target field characteristic model, the first object characteristic model and the first object pre-estimation model based on the difference value.
Optionally, the fifth training module 4021 includes: a sixth generating module 4022, configured to generate, by the cross-domain feature characterization model, second cross-domain feature characterization data of the sample user in the target domain based on the first feature data; a seventh generating module 4023, configured to generate, by the first object feature characterization model, feature characterization data of the sample object based on the fourth feature data; the second pre-estimation module 4024 is configured to pre-estimate, by using the first object pre-estimation model, the behavior of the sample user with respect to the sample object based on the second cross-domain feature characterization data and the feature characterization data of the sample object, so as to obtain behavior pre-estimation data of the sample user with respect to the sample object; an eighth training module 4025, configured to perform third training on the cross-domain feature characterization model, the first object feature characterization model, and the first object prediction model based on the behavior labeling data and the behavior prediction data.
Optionally, the eighth training module 4025 is specifically configured to: determining a difference value between the behavior marking data and the behavior estimation data through a target loss function; and adjusting parameters of the cross-domain feature characterization model, the first object feature characterization model and the first object pre-estimation model based on the difference values.
The user characterization device of this embodiment is used to implement the corresponding user characterization method in the foregoing multiple method embodiments, and has the beneficial effects of the corresponding method embodiments, which are not described herein again.
Referring to fig. 7, a schematic structural diagram of a user characterization device in the fifth embodiment of the present application is shown.
The user characterization device of the embodiment comprises: the first training module 501 is configured to train a cross-domain feature characterization model to be trained based on first feature data of a sample user in a source domain and second feature data of the sample user in a target domain; a first generating module 502, configured to generate, through the trained cross-domain feature characterization model, first cross-domain feature characterization data of the user to be characterized in the target domain based on third feature data of the user to be characterized in the source domain.
Optionally, the first generating module 502 includes: an eighth generating module 5021, configured to generate, through the trained cross-domain feature characterization model, third cross-domain feature characterization data of the to-be-characterized user for the object pre-estimation task in the target domain based on the third feature data of the to-be-characterized user in the source domain.
Optionally, the apparatus further comprises: a ninth generating module 503, configured to generate, by using a source domain feature characterization model, source domain feature characterization data of the to-be-characterized user for the object pre-estimation task based on third feature data of the to-be-characterized user in the source domain; a tenth generating module 504, configured to generate, by using a second target domain feature characterization model, second target domain feature characterization data of the to-be-characterized user for the object pre-estimation task based on fifth feature data of the to-be-characterized user in the target domain; a first determining module 505, configured to determine, based on the third cross-domain feature characterization data, the source domain feature characterization data, and the second target domain feature characterization data, comprehensive feature characterization data of the user to be characterized.
Optionally, the first determining module 505 includes: a first processing module 5051, configured to process, by using a domain attention model, the third cross-domain feature characterization data, the source domain feature characterization data, and the second target domain feature characterization data to obtain weight values corresponding to the third cross-domain feature characterization data, the source domain feature characterization data, and the second target domain feature characterization data, respectively; a second determining module 5052, configured to determine, based on weight values respectively corresponding to the third cross-domain feature characterization data, the source domain feature characterization data, and the second target domain feature characterization data, comprehensive feature characterization data of the user to be characterized is determined.
Optionally, the first processing module 5051 is specifically configured to: performing convolution operation on the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data through convolution layers of the domain attention model to obtain feature maps of the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data; performing mapping operation on the feature map through a full connection layer of the domain attention model to obtain weight feature vectors of the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data; and performing normalization operation on the weight feature vector through a calculation layer of the domain attention model to obtain weight values corresponding to the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data respectively.
Optionally, the apparatus further comprises: an eleventh generating module 506, configured to generate, by using a second object feature characterization model, feature characterization data of the object to be pre-estimated for the object pre-estimation task based on sixth feature data of the object to be pre-estimated in the target field; a third estimation module 507, configured to estimate, by using a second object estimation model, a behavior of the to-be-estimated user with respect to the to-be-estimated object based on the comprehensive characteristic characterization data of the to-be-estimated user and the characteristic characterization data of the to-be-estimated object with respect to the object estimation task, so as to obtain behavior estimation data of the to-be-estimated user with respect to the to-be-estimated object.
The user characterization device of this embodiment is used to implement the corresponding user characterization method in the foregoing multiple method embodiments, and has the beneficial effects of the corresponding method embodiments, which are not described herein again.
Fig. 8 is a schematic structural diagram of an electronic device in a sixth embodiment of the present application; the electronic device may include:
one or more processors 601;
a computer-readable medium 602, which may be configured to store one or more programs,
when executed by the one or more processors, cause the one or more processors to implement a user characterization method as described in the first or second embodiments above.
Fig. 9 is a hardware structure of an electronic device according to a seventh embodiment of the present application; as shown in fig. 9, the hardware structure of the electronic device may include: a processor 701, a communication interface 702, a computer-readable medium 703 and a communication bus 704;
wherein the processor 701, the communication interface 702, and the computer-readable medium 703 are in communication with each other via a communication bus 704;
alternatively, the communication interface 702 may be an interface of a communication module, such as an interface of a GSM module;
the processor 701 may be specifically configured to: training a cross-domain feature characterization model to be trained based on first feature data of a sample user in a source domain and second feature data of the sample user in a target domain; and generating first cross-domain feature characterization data of the user to be characterized in the target domain based on third feature data of the user to be characterized in the source domain through the trained cross-domain feature characterization model.
The Processor 701 may be a general-purpose Processor, and includes a Central Processing Unit (CPU), a Network Processor (NP), and the like; but may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), an off-the-shelf programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components. The various methods, steps, and logic blocks disclosed in the embodiments of the present application may be implemented or performed. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
The computer-readable medium 703 may be, but is not limited to, a Random Access Memory (RAM), a Read-Only Memory (ROM), a Programmable Read-Only Memory (PROM), an Erasable Read-Only Memory (EPROM), an electrically Erasable Read-Only Memory (EEPROM), and the like.
In particular, according to an embodiment of the present disclosure, the processes described above with reference to the flowcharts may be implemented as computer software programs. For example, embodiments of the present disclosure include a computer program product comprising a computer program embodied on a computer readable medium, the computer program comprising program code configured to perform the method illustrated by the flow chart. In such an embodiment, the computer program may be downloaded and installed from a network via the communication section, and/or installed from a removable medium. The computer program, when executed by a Central Processing Unit (CPU), performs the above-described functions defined in the method of the present application. It should be noted that the computer readable medium described herein can be a computer readable signal medium or a computer readable storage medium or any combination of the two. The computer readable medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. More specific examples of the computer readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access storage media (RAM), a read-only storage media (ROM), an erasable programmable read-only storage media (EPROM or flash memory), an optical fiber, a portable compact disc read-only storage media (CD-ROM), an optical storage media piece, a magnetic storage media piece, or any suitable combination of the foregoing. In the present application, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. In this application, however, a computer readable signal medium may include a propagated data signal with computer readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated data signal may take many forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A computer readable signal medium may also be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, fiber optic cable, RF, etc., or any suitable combination of the foregoing.
Computer program code configured to carry out operations for the present application may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C + +, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the case of a remote computer, the remote computer may operate over any of a variety of networks: including a Local Area Network (LAN) or a Wide Area Network (WAN) -to the user's computer, or alternatively, to an external computer (e.g., through the internet using an internet service provider).
The flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions configured to implement the specified logical function(s). In the above embodiments, specific precedence relationships are provided, but these precedence relationships are only exemplary, and in particular implementations, the steps may be fewer, more, or the execution order may be modified. That is, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.
The modules described in the embodiments of the present application may be implemented by software or hardware. The described modules may also be provided in a processor, which may be described as: a processor includes a first training module, a first generating module. The names of these modules do not constitute a limitation to the module itself in some cases, for example, the first training module may also be described as a "module for training a cross-domain feature characterization model to be trained based on first feature data of a sample user in a source domain and second feature data of the sample user in a target domain".
As another aspect, the present application also provides a computer-readable medium on which a computer program is stored, which when executed by a processor implements the user characterization method as described in the first or second embodiment.
As another aspect, the present application also provides a computer-readable medium, which may be contained in the apparatus described in the above embodiments; or may be present separately and not assembled into the device. The computer readable medium carries one or more programs which, when executed by the apparatus, cause the apparatus to: training a cross-domain feature characterization model to be trained based on first feature data of a sample user in a source domain and second feature data of the sample user in a target domain; and generating first cross-domain feature characterization data of the user to be characterized in the target domain based on third feature data of the user to be characterized in the source domain through the trained cross-domain feature characterization model.
The expressions "first", "second", "said first" or "said second" used in various embodiments of the present disclosure may modify various components regardless of order and/or importance, but these expressions do not limit the respective components. The above description is only configured for the purpose of distinguishing elements from other elements. For example, the first user equipment and the second user equipment represent different user equipment, although both are user equipment. For example, a first element could be termed a second element, and, similarly, a second element could be termed a first element, without departing from the scope of the present disclosure.
When an element (e.g., a first element) is referred to as being "operably or communicatively coupled" or "connected" (operably or communicatively) to "another element (e.g., a second element) or" connected "to another element (e.g., a second element), it is understood that the element is directly connected to the other element or the element is indirectly connected to the other element via yet another element (e.g., a third element). In contrast, it is understood that when an element (e.g., a first element) is referred to as being "directly connected" or "directly coupled" to another element (a second element), no element (e.g., a third element) is interposed therebetween.
The above description is only a preferred embodiment of the application and is illustrative of the principles of the technology employed. It will be appreciated by those skilled in the art that the scope of the invention herein disclosed is not limited to the particular combination of features described above, but also encompasses other arrangements formed by any combination of the above features or their equivalents without departing from the spirit of the invention. For example, the above features may be replaced with (but not limited to) features having similar functions disclosed in the present application.

Claims (30)

1. A method of user characterization, the method comprising:
training a cross-domain feature characterization model to be trained based on first feature data of a sample user in a source domain and second feature data of the sample user in a target domain;
and generating first cross-domain feature characterization data of the user to be characterized in the target domain based on third feature data of the user to be characterized in the source domain through the trained cross-domain feature characterization model.
2. The method of claim 1, wherein training the cross-domain feature characterization model to be trained based on the first feature data of the sample user in the source domain and the second feature data of the sample user in the target domain comprises:
training the cross-domain feature characterization model to be trained based on the first feature data, the second feature data, fourth feature data of a sample object associated with the sample user in the target domain, and behavior labeling data of the sample user for the sample object.
3. The method of claim 2, wherein the training the cross-domain feature characterization model to be trained based on the first feature data, the second feature data, fourth feature data of a sample object associated with the sample user in the target domain, and behavior labeling data of the sample user for the sample object comprises:
performing first training on the cross-domain feature characterization model and a first target domain feature characterization model based on the first feature data and the second feature data;
performing second training on the first target field characteristic representation model, the first object characteristic representation model and the first object estimation model based on the second characteristic data, the fourth characteristic data and the behavior marking data;
and performing third training on the cross-domain feature characterization model, the first object feature characterization model and the first object pre-estimation model based on the first feature data, the fourth feature data and the behavior labeling data.
4. The method of claim 3, wherein the first training of the cross-domain feature characterization model and the first target domain feature characterization model based on the first feature data and the second feature data comprises:
generating, by the cross-domain feature characterization model, second cross-domain feature characterization data of the sample user in the target domain based on the first feature data;
generating first target domain feature characterization data of the sample user based on the second feature data through the first target domain feature characterization model;
and performing first training on the cross-domain feature characterization model and the first target domain feature characterization model based on the second cross-domain feature characterization data and the first target domain feature characterization data.
5. The method of claim 4, wherein the first training of the cross-domain feature characterization model and the first target domain feature characterization model based on the second cross-domain feature characterization data and the first target domain feature characterization data comprises:
determining a distance value of the second cross-domain feature characterization data and the first target domain feature characterization data by a distance metric function;
adjusting parameters of the cross-domain feature characterization model and the first target domain feature characterization model based on the distance value.
6. The method of claim 3, wherein the second training of the first target domain feature characterization model, the first object feature characterization model and the first object pre-estimation model based on the second feature data, the fourth feature data and the behavior labeling data comprises:
generating first target domain feature characterization data of the sample user based on the second feature data through the first target domain feature characterization model;
generating, by the first object characterization model, characterization data for the sample object based on the fourth characterization data;
estimating the behavior of the sample user aiming at the sample object through the first object estimation model based on the first target field characteristic data and the characteristic data of the sample object so as to obtain behavior estimation data of the sample user aiming at the sample object;
and performing second training on the first target field characteristic representation model, the first object characteristic representation model and the first object estimation model based on the behavior marking data and the behavior estimation data.
7. The method of claim 6, wherein the second training of the first target domain feature characterization model, the first object feature characterization model, and the first object prediction model based on the behavior labeling data and the behavior prediction data comprises:
determining a difference value between the behavior marking data and the behavior estimation data through a target loss function;
and adjusting parameters of the first target field characteristic model, the first object characteristic model and the first object pre-estimation model based on the difference value.
8. The method of claim 3, wherein the third training of the cross-domain feature characterization model, the first object feature characterization model, and the first object prediction model based on the first feature data, the fourth feature data, and the behavior labeling data comprises:
generating, by the cross-domain feature characterization model, second cross-domain feature characterization data of the sample user in the target domain based on the first feature data;
generating, by the first object characterization model, characterization data for the sample object based on the fourth characterization data;
estimating the behavior of the sample user aiming at the sample object by the first object estimation model based on the second cross-domain feature characterization data and the feature characterization data of the sample object to obtain behavior estimation data of the sample user aiming at the sample object;
and performing third training on the cross-domain feature characterization model, the first object feature characterization model and the first object prediction model based on the behavior labeling data and the behavior prediction data.
9. The method of claim 8, wherein the third training of the cross-domain feature characterization model, the first object feature characterization model, and the first object prediction model based on the behavior labeling data and the behavior prediction data comprises:
determining a difference value between the behavior marking data and the behavior estimation data through a target loss function;
and adjusting parameters of the cross-domain feature characterization model, the first object feature characterization model and the first object pre-estimation model based on the difference values.
10. The method according to any one of claims 2 to 9, wherein the generating, by the trained cross-domain feature characterization model, first cross-domain feature characterization data of the user to be characterized in the target domain based on third feature data of the user to be characterized in the source domain comprises:
and generating third cross-domain feature characterization data of the user to be characterized aiming at the object pre-estimation task in the target domain based on the third feature data of the user to be characterized in the source domain through the trained cross-domain feature characterization model.
11. The method of claim 10, further comprising:
generating source field characteristic characterization data of the to-be-characterized user for the object pre-estimation task based on third characteristic data of the to-be-characterized user in the source field through a source field characteristic characterization model;
generating second target field characteristic representation data of the user to be characterized aiming at the object pre-estimation task on the basis of fifth characteristic data of the user to be characterized in the target field through a second target field characteristic representation model;
and determining comprehensive characteristic characterization data of the user to be characterized based on the third cross-domain characteristic characterization data, the source domain characteristic characterization data and the second target domain characteristic characterization data.
12. The method of claim 11, wherein determining the integrated feature characterization data for the user to be characterized based on the third cross-domain feature characterization data, the source domain feature characterization data, and the second target domain feature characterization data comprises:
processing the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data through a domain attention model to obtain weight values corresponding to the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data respectively;
determining comprehensive feature characterization data of the user to be characterized based on weighted values respectively corresponding to the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data, the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data.
13. The method according to claim 12, wherein the processing, by a domain attention model, the third cross-domain feature characterization data, the source domain feature characterization data, and the second target domain feature characterization data to obtain weight values corresponding to the third cross-domain feature characterization data, the source domain feature characterization data, and the second target domain feature characterization data, respectively, comprises:
performing convolution operation on the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data through convolution layers of the domain attention model to obtain feature maps of the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data;
performing mapping operation on the feature map through a full connection layer of the domain attention model to obtain weight feature vectors of the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data;
and performing normalization operation on the weight feature vector through a calculation layer of the domain attention model to obtain weight values corresponding to the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data respectively.
14. The method according to any one of claims 11-13, further comprising:
generating characteristic characterization data of the object to be pre-estimated aiming at the object pre-estimation task based on sixth characteristic data of the object to be pre-estimated in the target field through a second object characteristic characterization model;
estimating the behavior of the user to be characterized aiming at the object to be estimated through a second object estimation model based on the comprehensive characteristic characterization data of the user to be characterized and the characteristic characterization data of the object to be estimated aiming at the object estimation task, so as to obtain the behavior estimation data of the user to be characterized aiming at the object to be estimated.
15. A user characterization apparatus, the apparatus comprising:
the system comprises a first training module, a second training module and a third training module, wherein the first training module is used for training a cross-domain feature characterization model to be trained based on first feature data of a sample user in a source domain and second feature data of the sample user in a target domain;
the first generation module is used for generating first cross-domain feature characterization data of the user to be characterized in the target domain based on third feature data of the user to be characterized in the source domain through the trained cross-domain feature characterization model.
16. The apparatus of claim 15, wherein the first training module comprises:
and the second training module is used for training the cross-domain feature characterization model to be trained based on the first feature data, the second feature data, fourth feature data of a sample object associated with the sample user in the target domain and behavior marking data of the sample user for the sample object.
17. The apparatus of claim 16, wherein the second training module comprises:
a third training module, configured to perform first training on the cross-domain feature characterization model and the first target domain feature characterization model based on the first feature data and the second feature data;
the fourth training module is used for carrying out second training on the first target field characteristic representation model, the first object characteristic representation model and the first object estimation model based on the second characteristic data, the fourth characteristic data and the behavior marking data;
and the fifth training module is used for performing third training on the cross-domain feature characterization model, the first object feature characterization model and the first object pre-estimation model based on the first feature data, the fourth feature data and the behavior marking data.
18. The apparatus of claim 17, wherein the third training module comprises:
a second generation module, configured to generate, by the cross-domain feature characterization model, second cross-domain feature characterization data of the sample user in the target domain based on the first feature data;
a third generation module, configured to generate, by using the first target domain feature characterization model, first target domain feature characterization data of the sample user based on the second feature data;
a sixth training module, configured to perform first training on the cross-domain feature characterization model and the first target domain feature characterization model based on the second cross-domain feature characterization data and the first target domain feature characterization data.
19. The apparatus of claim 18, wherein the sixth training module is specifically configured to:
determining a distance value of the second cross-domain feature characterization data and the first target domain feature characterization data by a distance metric function;
adjusting parameters of the cross-domain feature characterization model and the first target domain feature characterization model based on the distance value.
20. The apparatus of claim 17, wherein the fourth training module comprises:
a fourth generation module, configured to generate, by using the first target domain feature characterization model, first target domain feature characterization data of the sample user based on the second feature data;
a fifth generation module, configured to generate, by the first object feature characterization model, feature characterization data of the sample object based on the fourth feature data;
the first estimation module is used for estimating the behavior of the sample user aiming at the sample object based on the first target field characteristic data and the characteristic data of the sample object through the first object estimation model so as to obtain behavior estimation data of the sample user aiming at the sample object;
and the seventh training module is used for carrying out second training on the first target field characteristic representation model, the first object characteristic representation model and the first object estimation model based on the behavior marking data and the behavior estimation data.
21. The apparatus of claim 20, wherein the seventh training module is specifically configured to:
determining a difference value between the behavior marking data and the behavior estimation data through a target loss function;
and adjusting parameters of the first target field characteristic model, the first object characteristic model and the first object pre-estimation model based on the difference value.
22. The apparatus of claim 17, wherein the fifth training module comprises:
a sixth generating module, configured to generate, by the cross-domain feature characterization model, second cross-domain feature characterization data of the sample user in the target domain based on the first feature data;
a seventh generating module, configured to generate, by the first object feature characterization model, feature characterization data of the sample object based on the fourth feature data;
the second estimation module is used for estimating the behavior of the sample user aiming at the sample object based on the second cross-domain feature characterization data and the feature characterization data of the sample object through the first object estimation model so as to obtain behavior estimation data of the sample user aiming at the sample object;
and the eighth training module is used for performing third training on the cross-domain feature characterization model, the first object feature characterization model and the first object prediction model based on the behavior marking data and the behavior prediction data.
23. The apparatus of claim 22, wherein the eighth training module is specifically configured to:
determining a difference value between the behavior marking data and the behavior estimation data through a target loss function;
and adjusting parameters of the cross-domain feature characterization model, the first object feature characterization model and the first object pre-estimation model based on the difference values.
24. The apparatus according to any one of claims 16-23, wherein the first generating module comprises:
and the eighth generation module is used for generating third cross-domain feature characterization data of the to-be-characterized user for the object pre-estimation task in the target domain based on the third feature data of the to-be-characterized user in the source domain through the trained cross-domain feature characterization model.
25. The apparatus of claim 24, further comprising:
a ninth generating module, configured to generate, through a source domain feature characterization model, source domain feature characterization data of the to-be-characterized user for the object pre-estimation task based on third feature data of the to-be-characterized user in the source domain;
a tenth generation module, configured to generate, by using a second target domain feature characterization model, second target domain feature characterization data of the to-be-characterized user for the object pre-estimation task based on fifth feature data of the to-be-characterized user in the target domain;
a first determining module, configured to determine, based on the third cross-domain feature characterization data, the source domain feature characterization data, and the second target domain feature characterization data, comprehensive feature characterization data of the user to be characterized.
26. The apparatus of claim 25, wherein the first determining module comprises:
a first processing module, configured to process, through a domain attention model, the third cross-domain feature characterization data, the source domain feature characterization data, and the second target domain feature characterization data to obtain weight values corresponding to the third cross-domain feature characterization data, the source domain feature characterization data, and the second target domain feature characterization data, respectively;
a second determining module, configured to determine comprehensive feature characterization data of the user to be characterized based on weight values corresponding to the third cross-domain feature characterization data, the source domain feature characterization data, and the second target domain feature characterization data, respectively, the third cross-domain feature characterization data, the source domain feature characterization data, and the second target domain feature characterization data.
27. The apparatus of claim 26, wherein the first processing module is specifically configured to:
performing convolution operation on the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data through convolution layers of the domain attention model to obtain feature maps of the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data;
performing mapping operation on the feature map through a full connection layer of the domain attention model to obtain weight feature vectors of the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data;
and performing normalization operation on the weight feature vector through a calculation layer of the domain attention model to obtain weight values corresponding to the third cross-domain feature characterization data, the source domain feature characterization data and the second target domain feature characterization data respectively.
28. The apparatus according to any one of claims 25-27, further comprising:
an eleventh generating module, configured to generate, through a second object feature characterization model, feature characterization data of the object to be pre-estimated for the object pre-estimation task based on sixth feature data of the object to be pre-estimated in the target field;
and the third estimation module is used for estimating the behavior of the user to be characterized aiming at the object to be estimated through a second object estimation model based on the comprehensive characteristic representation data of the user to be characterized and the characteristic representation data of the object to be estimated aiming at the object estimation task so as to obtain the behavior estimation data of the user to be characterized aiming at the object to be estimated.
29. An electronic device, comprising:
one or more processors;
a computer readable medium configured to store one or more programs,
the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the user characterization method of any one of claims 1-14.
30. A computer-readable medium, on which a computer program is stored which, when being executed by a processor, carries out the user characterization method according to any one of claims 1-14.
CN201910131144.3A 2019-02-21 2019-02-21 User characterization method, device, electronic equipment and computer readable medium Active CN111598599B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910131144.3A CN111598599B (en) 2019-02-21 2019-02-21 User characterization method, device, electronic equipment and computer readable medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910131144.3A CN111598599B (en) 2019-02-21 2019-02-21 User characterization method, device, electronic equipment and computer readable medium

Publications (2)

Publication Number Publication Date
CN111598599A true CN111598599A (en) 2020-08-28
CN111598599B CN111598599B (en) 2023-04-25

Family

ID=72181497

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910131144.3A Active CN111598599B (en) 2019-02-21 2019-02-21 User characterization method, device, electronic equipment and computer readable medium

Country Status (1)

Country Link
CN (1) CN111598599B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112541132A (en) * 2020-12-23 2021-03-23 北京交通大学 Cross-domain recommendation method based on multi-view knowledge representation

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140019450A1 (en) * 2012-07-12 2014-01-16 Daniela Rus Text characterization of trajectories
CN106649434A (en) * 2016-09-06 2017-05-10 北京蓝色光标品牌管理顾问股份有限公司 Cross-domain knowledge transfer tag embedding method and apparatus
CN107291840A (en) * 2017-05-31 2017-10-24 北京奇艺世纪科技有限公司 A kind of user property forecast model construction method and device
CN108898218A (en) * 2018-05-24 2018-11-27 阿里巴巴集团控股有限公司 A kind of training method of neural network model, device and computer equipment
CN109345302A (en) * 2018-09-27 2019-02-15 腾讯科技（深圳）有限公司 Machine learning model training method, device, storage medium and computer equipment

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140019450A1 (en) * 2012-07-12 2014-01-16 Daniela Rus Text characterization of trajectories
CN106649434A (en) * 2016-09-06 2017-05-10 北京蓝色光标品牌管理顾问股份有限公司 Cross-domain knowledge transfer tag embedding method and apparatus
CN107291840A (en) * 2017-05-31 2017-10-24 北京奇艺世纪科技有限公司 A kind of user property forecast model construction method and device
CN108898218A (en) * 2018-05-24 2018-11-27 阿里巴巴集团控股有限公司 A kind of training method of neural network model, device and computer equipment
CN109345302A (en) * 2018-09-27 2019-02-15 腾讯科技（深圳）有限公司 Machine learning model training method, device, storage medium and computer equipment

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
赵传君;王素格;李德玉;: "基于集成深度迁移学习的多源跨领域情感分类" *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112541132A (en) * 2020-12-23 2021-03-23 北京交通大学 Cross-domain recommendation method based on multi-view knowledge representation
CN112541132B (en) * 2020-12-23 2023-11-10 北京交通大学 Cross-domain recommendation method based on multi-view knowledge representation

Also Published As

Publication number Publication date
CN111598599B (en) 2023-04-25

Similar Documents

Publication Publication Date Title
WO2020156389A1 (en) Information pushing method and device
CN111104599B (en) Method and device for outputting information
CN112650841A (en) Information processing method and device and electronic equipment
CN114117216A (en) Recommendation probability prediction method and device, computer storage medium and electronic equipment
CN113610582A (en) Advertisement recommendation method and device, storage medium and electronic equipment
CN114417174B (en) Content recommendation method, device, equipment and computer storage medium
CN111444335B (en) Method and device for extracting central word
CN115631012A (en) Target recommendation method and device
CN115018552A (en) Method for determining click rate of product
US20220358366A1 (en) Generation and implementation of dedicated feature-based techniques to optimize inference performance in neural networks
CN111507471B (en) Model training method, device, equipment and storage medium
CN111598599B (en) User characterization method, device, electronic equipment and computer readable medium
CN116910373A (en) House source recommendation method and device, electronic equipment and storage medium
WO2023185125A1 (en) Product resource data processing method and apparatus, electronic device and storage medium
CN110930226A (en) Financial product recommendation method and device, electronic equipment and storage medium
CN114625876B (en) Method for generating author characteristic model, method and device for processing author information
CN113723712B (en) Wind power prediction method, system, equipment and medium
CN113837808B (en) Promotion information pushing method, device, equipment, medium and product
CN116109374A (en) Resource bit display method, device, electronic equipment and computer readable medium
CN115391661A (en) Method for training recommendation model, method for recommending resources and device thereof
CN111581455B (en) Text generation model generation method and device and electronic equipment
CN113609397A (en) Method and device for pushing information
CN114092162A (en) Recommendation quality determination method, and training method and device of recommendation quality determination model
CN111784377A (en) Method and apparatus for generating information
CN116911913B (en) Method and device for predicting interaction result

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant