CN115730125A

CN115730125A - Object identification method and device, computer equipment and storage medium

Info

Publication number: CN115730125A
Application number: CN202110983637.7A
Authority: CN
Inventors: 乔阳; 陈亮
Original assignee: Tenpay Payment Technology Co Ltd
Current assignee: Tenpay Payment Technology Co Ltd
Priority date: 2021-08-25
Filing date: 2021-08-25
Publication date: 2023-03-03

Abstract

The application relates to an object identification method, an object identification device, a computer device and a storage medium. The method comprises the following steps: inputting the training sample into a shared feature extraction network in an object recognition model for feature extraction to obtain sample extraction features, and inputting the sample extraction features of which the target sample type is a push sample type into a push object recognition network in the object recognition model to obtain the content conversion degree of a push object; inputting the sample extraction characteristics of which the target sample type is the reference sample type into a reference object identification network in an object identification model to obtain the content conversion degree of the reference object; obtaining a pushing identification loss value based on the content conversion degree of the pushing object, and obtaining a reference identification loss value based on the content degree of the reference object; and performing parameter adjustment on the shared feature extraction network based on the pushing recognition loss value and the reference recognition loss value to obtain a trained object recognition model. By adopting the method, the accuracy of object identification can be improved.

Description

Object identification method and device, computer equipment and storage medium

Technical Field

The present application relates to the field of computer technologies, and in particular, to an object identification method and apparatus, a computer device, and a storage medium.

Background

With the development of computer technology and internet technology, more and more contents are pushed to users through the internet, for example, advertisements are pushed to users or coupons are issued to users.

In many cases, object recognition is often required to identify the object from which the content is pushed. For example, since different user groups may react differently to the same push content, it is necessary to identify a user matching the push content, so as to push the content to this part of users, so as to improve the efficiency of content push.

Disclosure of Invention

In view of the above, it is necessary to provide an object recognition method, an apparatus, a computer device, and a storage medium capable of improving object recognition accuracy in view of the above technical problems.

An object recognition method, the method comprising: acquiring a training sample set corresponding to an object recognition model to be trained, wherein the training sample set comprises training samples corresponding to content push objects and training samples corresponding to reference objects; respectively inputting the training samples in the training sample set into a shared feature extraction network in the object recognition model for feature extraction to obtain sample extraction features corresponding to the training samples; determining a target sample type corresponding to the sample extraction features; inputting the sample extraction characteristics of which the target sample type is the push sample type into a push object identification network in the object identification model for identification, and identifying to obtain the content conversion degree of a push object corresponding to the content push object; inputting sample extraction features of which the target sample type is a reference sample type into a reference object identification network in the object identification model for identification, and identifying to obtain a reference object content conversion degree corresponding to the reference object; obtaining a pushing identification loss value based on a pushing object content conversion degree corresponding to the content pushing object, and obtaining a reference identification loss value based on a reference object content conversion degree corresponding to the reference object; and performing parameter adjustment on the shared feature extraction network based on the push identification loss value and the reference identification loss value, and obtaining a trained object identification model based on the adjusted shared feature extraction network.

An object recognition apparatus, the apparatus comprising: the training sample set acquisition module is used for acquiring a training sample set corresponding to an object recognition model to be trained, wherein the training sample set comprises a training sample corresponding to a content push object and a training sample corresponding to a reference object; a sample extraction feature obtaining module, configured to input training samples in the training sample set to a shared feature extraction network in the object recognition model respectively for feature extraction, so as to obtain sample extraction features corresponding to the training samples; the target sample type determining module is used for determining a target sample type corresponding to the sample extraction features; a pushed object content conversion degree obtaining module, configured to input the sample extraction feature that the target sample type is the pushed sample type into a pushed object identification network in the object identification model for identification, and identify to obtain a pushed object content conversion degree corresponding to the content pushed object; a reference object content conversion degree obtaining module, configured to input a sample extraction feature that a target sample type is a reference sample type into a reference object identification network in the object identification model for identification, and obtain a reference object content conversion degree corresponding to the reference object through identification; an identification loss value obtaining module, configured to obtain a push identification loss value based on a push object content conversion degree corresponding to the content push object, and obtain a reference identification loss value based on a reference object content conversion degree corresponding to the reference object; and the trained object recognition model obtaining module is used for carrying out parameter adjustment on the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and obtaining a trained object recognition model based on the adjusted shared feature extraction network.

A computer device comprising a memory storing a computer program and a processor implementing the steps of the object recognition method described above when executing the computer program.

A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of the above-mentioned object recognition method.

According to the object identification method, the device, the computer equipment and the storage medium, the sample extraction features corresponding to the training samples of the content push objects are obtained by extracting the shared feature extraction network, and the sample extraction features corresponding to the training samples of the reference objects are also obtained by extracting the shared feature extraction network, so that the training samples of the content push objects and the training samples of the reference objects share one shared feature extraction network, and under the condition that the number of the samples is unbalanced, the feature extraction network can be trained by adopting two different samples, namely the training samples of the content push objects and the training samples of the reference objects, so that the shared feature extraction network learns the capacity of distinguishing the two samples, the model training effect is improved, the accuracy of the trained object identification model is improved, and the accuracy of the object identification is improved.

A method of object recognition, the method comprising: acquiring a candidate object set corresponding to target push content; the candidate object set comprises a plurality of candidate objects; inputting the object sample of the candidate object into a shared feature extraction network of a trained object recognition model for feature extraction to obtain object extraction features corresponding to the candidate object; inputting the object extraction features into a pushed object recognition network in the object recognition model for recognition, and obtaining a first content conversion possibility corresponding to the candidate object based on the conversion possibility obtained through recognition; inputting the object extraction features into a reference object recognition network in the object recognition model for recognition, and obtaining a second content conversion possibility corresponding to the candidate object based on the conversion possibility obtained through recognition; acquiring a possibility difference between a first content conversion possibility and a second content conversion possibility corresponding to the candidate object; screening candidate objects meeting difference conditions from the candidate object set based on the difference of the possibility degrees corresponding to the candidate objects to serve as push objects corresponding to the target push content; the difference condition includes at least one of a ranking of the energy difference before a preset ranking or a likelihood difference greater than a difference threshold.

An object recognition apparatus, the apparatus comprising: the candidate object set acquisition module is used for acquiring a candidate object set corresponding to the target push content; the candidate object set comprises a plurality of candidate objects; an object extraction feature obtaining module, configured to input the object sample of the candidate object into a shared feature extraction network of a trained object recognition model to perform feature extraction, so as to obtain an object extraction feature corresponding to the candidate object; a first content conversion possibility obtaining module, configured to input the object extraction feature into a pushed object recognition network in the object recognition model for recognition, and obtain a first content conversion possibility corresponding to the candidate object based on the conversion possibility obtained through recognition; a second content conversion possibility obtaining module, configured to input the object extraction feature into a reference object recognition network in the object recognition model for recognition, and obtain a second content conversion possibility corresponding to the candidate object based on the conversion possibility obtained through recognition; a likelihood difference obtaining module, configured to obtain a likelihood difference between a first content conversion likelihood and a second content conversion likelihood corresponding to the candidate object; a pushed object determining module, configured to filter candidate objects that meet a difference condition from the candidate object set based on a difference in likelihood corresponding to the candidate objects, and use the candidate objects as pushed objects corresponding to the target pushed content; the difference condition includes at least one of a ranking of the energy difference before a preset ranking or a likelihood difference greater than a difference threshold.

In some embodiments, a computer program product or computer program is provided that includes computer instructions stored in a computer-readable storage medium. The computer instructions are read by a processor of a computer device from a computer-readable storage medium, and the computer instructions are executed by the processor to cause the computer device to perform the steps in the above-mentioned method embodiments.

The object identification method, the device, the computer equipment and the storage medium obtain a candidate object set corresponding to target pushed content, the candidate object set comprises a plurality of candidate objects, an object sample of the candidate objects is input into a shared feature extraction network of a trained object identification model for feature extraction to obtain object extraction features corresponding to the candidate objects, the object extraction features are input into a pushed object identification network of the object identification model for identification, first content conversion possibility corresponding to the candidate objects is obtained based on the conversion possibility obtained through identification, the object extraction features are input into a reference object identification network of the object identification model for identification, second content conversion possibility corresponding to the candidate objects is obtained based on the conversion possibility obtained through identification, a possibility difference between the first content conversion possibility and the second content conversion possibility corresponding to the candidate objects is obtained, the candidate objects meeting a difference condition are screened from the candidate object set and serve as the pushed objects corresponding to the target pushed content, the condition comprises that the ranking of the possibility difference is before presetting or the ranking of the possibility difference is greater than a threshold value, and at least one type of the corresponding pushed objects is convenient for identification of the target objects.

Drawings

FIG. 1 is a diagram of an application environment of an object recognition method in some embodiments;

FIG. 2 is a flow diagram that illustrates a method for object recognition in some embodiments;

FIG. 3 is a model structure diagram of a delayed feedback model in some embodiments;

FIG. 4 is a block diagram of an object recognition model in some embodiments;

FIG. 5 is a flow diagram that illustrates a method for object recognition in some embodiments;

FIG. 6 is a block diagram of a process for training two models in some embodiments;

FIG. 7 is a block diagram of a process for gain value prediction using two trained models in some embodiments;

FIG. 8 is a model structure diagram of a single model in some embodiments;

FIG. 9 is a block diagram of a model architecture of a class transformation method in some embodiments;

FIG. 10 is a block diagram of the structure of an object recognition device in some embodiments;

FIG. 11 is a block diagram of an object recognition device in some embodiments;

FIG. 12 is a diagram of the internal structure of a computer device in some embodiments;

FIG. 13 is a diagram of the internal structure of a computer device in some embodiments.

Detailed Description

In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.

Artificial Intelligence (AI) is a theory, method, technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human Intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a manner similar to human intelligence. Artificial intelligence is the research of the design principle and the realization method of various intelligent machines, so that the machines have the functions of perception, reasoning and decision making.

The artificial intelligence technology is a comprehensive subject and relates to the field of extensive technology, namely the technology of a hardware level and the technology of a software level. The artificial intelligence base technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning, automatic driving, intelligent traffic and the like.

Computer Vision technology (CV) Computer Vision is a science for researching how to make a machine "see", and further refers to that a camera and a Computer are used to replace human eyes to perform machine Vision such as identification, tracking and measurement on a target, and further image processing is performed, so that the Computer processing becomes an image more suitable for human eyes to observe or transmitted to an instrument to detect. As a scientific discipline, computer vision research-related theories and techniques attempt to build artificial intelligence systems that can capture information from images or multidimensional data. The computer vision technology generally includes image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content/behavior recognition, three-dimensional object reconstruction, 3D technology, virtual reality, augmented reality, synchronous positioning and map construction, automatic driving, intelligent transportation and other technologies, and also includes common biometric identification technologies such as face recognition and fingerprint recognition.

Machine Learning (ML) is a multi-domain cross subject, and relates to multi-domain subjects such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and the like. The special research on how a computer simulates or realizes the learning behavior of human beings so as to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve the performance of the computer. Machine learning is the core of artificial intelligence, is the fundamental approach for computers to have intelligence, and is applied to all fields of artificial intelligence. Machine learning and deep learning generally include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and formula learning.

With the research and development of artificial intelligence technology, the artificial intelligence technology is developed and researched in a plurality of fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, automatic driving, unmanned aerial vehicles, robots, smart medical services, smart customer service, internet of vehicles, automatic driving, smart transportation and the like.

The scheme provided by the embodiment of the application relates to the technologies such as machine learning of artificial intelligence and the like, and is specifically explained by the following embodiment:

the object identification method provided by the application can be applied to the application environment shown in fig. 1. Wherein the terminal 102 communicates with the server 104 via a network.

The server 104 may be a server for content push, for example, a server for pushing content into an Application (Application), for example, a server for pushing advertisements, and the Application may include at least one of an instant messaging Application or a shopping Application. The terminal 102 may be installed with a client corresponding to the application program, the server 104 may push the content to the client corresponding to the application program in the terminal 102, and the terminal 102 may display the pushed content in the client, for example, the client may display the pushed content in a pushed content display area, and when the pushed content is an advertisement, the pushed content display area may be an advertisement space.

Specifically, the server 104 may obtain a training sample set corresponding to an object recognition model to be trained, where the training sample set includes training samples corresponding to content push objects and training samples corresponding to reference objects, respectively input the training samples in the training sample set into a shared feature extraction network in the object recognition model to perform feature extraction, obtain sample extraction features corresponding to the training samples, determine a target sample type corresponding to the sample extraction features, input the sample extraction features of which the target sample type is the push sample type into a push object recognition network in the object recognition model to perform recognition, obtain a push object content conversion degree of the content push objects, input the sample extraction features of which the target sample type is the reference sample type into a reference object recognition network in the object recognition model to perform recognition, obtain a reference object content conversion degree of the reference objects, obtain a push recognition loss value based on the push object content conversion degree, obtain a reference recognition loss value based on the reference object content conversion degree, perform parameter adjustment on the push recognition loss value and the reference recognition loss value, and perform parameter adjustment on the shared feature extraction network after adjustment, and obtain an adjusted shared object recognition model. The content push object refers to an object to which training push content is pushed, and the reference object refers to an object to which training push content is not pushed.

The server 104 may obtain a candidate object set in response to a content push request for a target push content, where the candidate object set includes a plurality of candidate objects, input an object sample of the candidate objects into a shared feature extraction network of a trained object recognition model to perform feature extraction, obtain object extraction features corresponding to the candidate objects, input the object extraction features into a pushed object recognition network of the object recognition model to perform recognition, obtain a first content conversion possibility corresponding to the candidate objects based on the conversion possibility obtained through recognition, input the object extraction features into a reference object recognition network of the object recognition model to perform recognition, obtain a second content conversion possibility corresponding to the candidate objects based on the conversion possibility obtained through recognition, obtain a possibility difference between the first content conversion possibility and the second content conversion possibility corresponding to the candidate objects, and screen candidate objects satisfying a difference condition from the candidate object set based on the possibility difference corresponding to obtain candidate objects as push objects corresponding to the target push content; the difference condition includes at least one of a ranking of the energy difference before a preset ranking or a likelihood difference greater than a difference threshold. The server 104 may push the targeted push content to the push object. The target push content may be the same as the training push content, and the target push content may also be different from the training push content.

The terminal 102 may be, but not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices, and the server 104 may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing basic cloud computing services such as cloud storage, network services, cloud communication, big data, and an artificial intelligence platform. The terminal and the server may be directly or indirectly connected through wired or wireless communication, and the application is not limited herein.

It may be understood that the foregoing application scenario is only an example, and does not constitute a limitation on the object recognition provided in the embodiment of the present application, and the method provided in the embodiment of the present application may also be applied in other application scenarios, for example, the object recognition provided in the present application may be executed by the terminal 102, the terminal 102 may upload the obtained trained object recognition model to the server 104, the server 104 may store the trained object recognition model, and may also forward the trained object recognition model to other devices.

The object recognition method provided by the present application, wherein the server 104 may be a node in the blockchain, and the server 104 may store the trained object recognition model in the blockchain. The server 104 may further filter a push object corresponding to the push content from the candidate object set, establish a correspondence between the push content and the push object, store the correspondence between the push content and the push object in the blockchain, and the server 104 may obtain the push object corresponding to the push content by querying from the blockchain according to the push content and push the push content to the push object.

In some embodiments, as shown in fig. 2, an object identification method is provided, where the method may be executed by a terminal or a server, or may be executed by both the terminal and the server, and in this embodiment, the method is described as applied to the server 104 in fig. 1, and includes the following steps:

s202, a training sample set corresponding to an object recognition model to be trained is obtained, wherein the training sample set comprises training samples corresponding to content pushing objects and training samples corresponding to reference objects.

The object may be a natural person, for example, a user using an application program, and the application program may include a shopping application program or a financial application program. The content push object is an object to which content is pushed, the content may include virtual things that can be pushed through a network, for example, an electronic coupon or an advertisement, and the content may also include real things, for example, a coupon in a paper form. The reference object is a concept opposite to the content push object, the reference object refers to an object to which content is not pushed, and the content pushed to the content push object and the content not pushed to the reference object may be the same content.

The conversion is determined according to the purpose of the pushed content, when the purpose of the pushed content is to stimulate the user to click, if the object clicks the pushed content, the object is determined to be converted, and when the purpose of the content pushing is to stimulate the user to purchase goods related to the pushed content, if the object purchases the goods related to the pushed content, the object is determined to be converted. The content push object may also be referred to as an intervention object and the reference object may also be referred to as a reference object.

The training sample set comprises a plurality of training samples, and the training samples are used for training the object recognition model. The training sample may be determined according to the object attribute information, for example, the object attribute information may be used as the training sample, and the object attribute information may include at least one of age, occupation, academic calendar, hobbies, located area, or gender of the object. The training sample of the content pushing object can be determined according to the object attribute information corresponding to the content pushing object, and the training of the reference object can be determined according to the object attribute information corresponding to the reference object. The training samples of the content-pushed object may be referred to as intervention samples, and the training samples corresponding to the reference object may be referred to as control samples. Where intervention (transaction) is used to indicate that content push is performed, for example, issuing a coupon to an object is an intervention and pushing an advertisement is an intervention. The training samples corresponding to the content push objects may also be referred to as push training samples, and the training samples corresponding to the reference objects may also be referred to as reference training samples.

The training samples may include positive samples and may also include negative samples, the training samples of the content push objects may include positive samples and negative samples, and the training samples corresponding to the reference objects may include positive samples and negative samples.

The positive sample refers to a training sample corresponding to an object which has undergone conversion within the observation time period, and the negative sample refers to a training sample corresponding to an object which has not undergone conversion within the observation time period. The observation time period is a time period from the content push time to the positive and negative sample division reference time, and the positive and negative sample division reference time may also be referred to as an observation day. The content push time is a time when the content is pushed to the content push object. Since the same object has not been converted in the previous time, but may be converted in the next time, that is, the conversion result of the same object may be different in different times, that is, the conversion result varies with time, and the conversion result may be either converted or unconverted, it is necessary to determine whether the training sample is a positive sample or a negative sample according to time, where the positive and negative sample division reference time refers to a time referenced for dividing the training sample into a positive sample or a negative sample, for example, the content push time is 1 month and 1 day, and the positive and negative sample division reference time is 1 month and 10 days, and the observation time period is a time period from 1 month and 1 day to 1 month and 10 days, if the object has been converted in the time period from 1 month and 1 day to 1 month and 10 days, the training sample corresponding to the object is a positive sample, and if the object has not been converted in the time period from 1 month and 1 day to 1 month and 10 days, the training sample corresponding to the object is a negative sample. The positive and negative sample division reference time may be preset or set as needed.

The number of the positive samples and the number of the negative samples may be the same or different, for example, in a financial scenario, since the reference object is not interfered, the conversion of the reference object depends on the natural conversion of the object, and the conversion rate is low, so that the number of the positive samples corresponding to the reference object is smaller than the number of the negative samples corresponding to the reference object.

The object recognition model may be an artificial intelligence based neural network model. The object recognition model is used to identify object types, which may include at least one of marketing sensitive objects, natural conversion objects, immoderate objects, or counteractive objects. The object recognition model to be trained may be an untrained model or a model that is trained but needs to be further trained. The marketing sensitive object is an object which does not convert when the content is not pushed, and converts when the content is pushed, the natural conversion object is an object which can convert whether the content is pushed or not, the involuntary object is an object which cannot convert whether the content is pushed or not, the counteractive object is an object which is relatively insensitive to marketing activities, can convert when the content is not pushed, and cannot convert when the content is pushed. For example, the marketing sensitive object refers to an object which is not purchased without issuing a ticket and is purchased only after issuing a ticket, the natural conversion object refers to an object which is purchased regardless of whether the ticket is issued or not, the involuntary object refers to an object which is not purchased regardless of whether the ticket is issued or not, the counteractive object refers to an object which is relatively insensitive to marketing activities and is purchased when the ticket is not issued but is not purchased after issuing the ticket. The marketing sensitive object can be identified through the object identification model, so that the content can be pushed to the marketing sensitive object, the marketing sensitive object is stimulated to be converted, the content pushing accuracy is improved, and the content pushing cost is reduced.

Specifically, the server may obtain an object set, where the object set includes a plurality of objects, determine training push content to be pushed, where the training push content may be any content, randomly select a plurality of first objects from the object set, use objects in the object set other than the first objects as second objects, push the training push content to each first object, use the first object from which the training push content is pushed as a content push object, and use each second object as a reference object. The number of the first objects and the number of the second objects may be the same or different, that is, the number of the pushed content objects and the number of the reference objects may be the same or different.

In some embodiments, the number of training samples corresponding to the content push object is less than the number of training samples corresponding to the reference object, for example, in a financial scenario, it is difficult to reserve more reference objects due to low object conversion rate and low object behavior frequency, the reference object may also be referred to as an unbiased comparison object, and the training samples corresponding to the reference object may also be referred to as unbiased comparison samples, that is, the unbiased comparison samples are scarce in the financial scenario, and the user has high decision cost and is difficult to convert, so that the number of samples of the content push object is less than the number of samples of the reference object.

In some embodiments, the number of training samples corresponding to the content push object is less than the number of training samples corresponding to the reference object, and the number of positive samples in the training samples corresponding to the reference object is less than the number of negative samples in the training samples corresponding to the reference object.

And S204, respectively inputting the training samples in the training sample set into a shared feature extraction network in the object recognition model for feature extraction, and obtaining sample extraction features corresponding to the training samples.

The object recognition model may include a shared feature extraction network, and the sample extraction features are features obtained by performing feature extraction on training samples by the shared feature extraction network.

Specifically, the server may input the training sample corresponding to the content push object into the shared feature extraction network for feature extraction to obtain the sample extraction feature of the training sample corresponding to the content push object, and may input the training sample corresponding to the reference object into the shared feature extraction network for feature extraction to obtain the sample extraction feature of the training sample corresponding to the reference object. When training the object recognition model, one training sample may be input into the object recognition model at a time, that is, one training sample is used for one training, but a plurality of training samples may be used for one training, for example, all training samples may be used.

In some embodiments, the shared feature extraction network may include an Embedding layer (Embedding), an aggregation layer (aggregating), and a feature extraction layer, the training sample includes a plurality of object attribute information, for example, age, occupation, gender, and the like, the server may perform word vector conversion using each object attribute information in the embedded layer training sample in the shared feature extraction network to obtain word vectors corresponding to each object attribute information, input the word vectors corresponding to each object attribute information to the aggregation layer in the shared feature extraction network to perform aggregation to obtain an aggregation result, and input the aggregation result to the feature extraction layer to perform further feature extraction to obtain sample extraction features corresponding to the training sample. The aggregation may be obtained by splicing word vectors of each object attribute information, or by performing weighted calculation on word vectors of each object attribute information. The feature extraction layer may include at least one of a convolutional layer or an active layer.

And S206, determining the target sample type corresponding to the sample extraction features.

The sample type is used for distinguishing whether the training sample is a training sample corresponding to a content push object or a training sample corresponding to a reference object, the sample type may include a push sample type and a reference sample type, and the target sample type may be any one of the push sample type and the reference sample type. The target sample type corresponding to the sample extraction feature refers to a sample type to which a training sample corresponding to the sample extraction feature belongs, if the target sample type is a push sample type, the sample extraction feature is a feature obtained by performing feature extraction on the training sample of the content push object, and if the target sample type is a reference sample type, the sample extraction feature is a feature obtained by performing feature extraction on the training sample of the reference object.

Specifically, when the server obtains the sample extraction features corresponding to the training samples, the sample types corresponding to the sample extraction features may be determined, for example, the sample types may be determined according to sample label values of the training samples, where the sample label values are used to distinguish the sample types, and the sample label values are different and the corresponding sample types are different. The sample flag value may be preset or set as needed, for example, the sample flag value corresponding to the push sample type may be set to 1, and the sample flag value corresponding to the reference sample type may be set to 0.

And S208, inputting the sample extraction characteristics of which the target sample type is the push sample type into a push object identification network in the object identification model for identification, and identifying to obtain the content conversion degree of the push object corresponding to the content push object.

The object recognition model can also comprise a push object recognition network. The push object recognition network inputs sample extraction characteristics of training samples corresponding to the content push objects.

If the training sample corresponding to the content push object is a positive sample, the push object content conversion degree is a possibility that the content push object is converted at the positive and negative sample division reference time, for example, a possibility that the content push object clicks the push content at the positive and negative sample division reference time, or a possibility that the content push object purchases a commodity related to the pushed content at the positive and negative sample division reference time, for example, if the pushed content is a coupon of a computer, the conversion possibility may be a possibility that the user receiving the coupon purchases the computer. For example, if the observation time period is 1 month and 1 day to 1 month and 10 days, the positive and negative sample division reference time is 1 month and 10 days, and the push object content conversion degree is used to reflect the possibility that the content push object is converted in 1 month and 10 days. The content conversion degree of the push object is in positive correlation with the possibility that the content push object is converted in the positive and negative sample division reference time.

If the training sample corresponding to the content push object is a negative sample, the content conversion degree of the push object refers to the possibility that the content push object is not converted in the positive and negative sample division reference time. The content conversion degree of the push object is in positive correlation with the possibility that the content push object does not convert in the positive and negative sample division reference time.

The positive correlation refers to: under the condition that other conditions are not changed, the changing directions of the two variables are the same, and when one variable is changed from large to small, the other variable is also changed from large to small. It is understood that a positive correlation herein means that the direction of change is consistent, but does not require that when one variable changes at all, another variable must also change. For example, it may be set that the variable b is 100 when the variable a is 10 to 20, and the variable b is 120 when the variable a is 20 to 30. Thus, the change directions of a and b are both such that when a is larger, b is also larger. But b may be unchanged in the range of 10 to 20 a.

Specifically, the server may determine a sample type corresponding to the sample extraction feature, and if the sample type corresponding to the sample extraction feature is a push sample type, input the sample extraction feature into a push object identification network for identification.

In some embodiments, the pushed object identification network may include a conversion likelihood identification network and a likelihood attenuation factor identification network, the server may input the sample extraction feature into the conversion likelihood identification network to identify the conversion likelihood, obtain a total conversion likelihood of the content pushed object, input the sample extraction feature into the likelihood attenuation factor identification network to identify the attenuation factor, obtain a likelihood attenuation factor corresponding to the content pushed object, and determine the sub-conversion likelihood of the content pushed object based on the likelihood attenuation factor, where the likelihood attenuation factor is a parameter of a distribution function to which the likelihood of the conversion of the object decays with time, that is, a parameter of a distribution function to which the decay of the likelihood of the conversion of the object changes with time changes, for example, when the decay of the likelihood of the conversion of the object changes with time obeys an exponential distribution, the distribution function may be an exponential distribution function, and the sub-conversion likelihood refers to a likelihood that the object to be finally converted is converted at a certain time, for example, a probability that the object to be finally converted is converted on the day d after the content is pushed. The total conversion possibility degree refers to a possibility that the object will be converted finally, for example, a probability that a user will purchase a commodity in the pushed advertisement finally, and the server may calculate the pushed object content conversion degree of the content pushing object according to the total conversion possibility degree of the content pushing object and the sub-conversion possibility degree corresponding to the content pushing object.

The first conversion possibility degree identification network mentioned below refers to a conversion possibility degree identification network in the push object identification network, and the first possibility degree attenuation factor identification network refers to a possibility degree attenuation factor identification network in the push object identification network.

And S210, inputting the sample extraction characteristics of which the target sample type is the reference sample type into a reference object identification network in the object identification model for identification, and identifying to obtain the content conversion degree of the reference object corresponding to the reference object.

Wherein, the object recognition model can also comprise a reference object recognition network. The reference object recognition network inputs sample extraction features of training samples corresponding to the reference objects. If the training sample corresponding to the reference object is a positive sample, the reference object content conversion degree refers to the possibility that the reference object is converted in the positive and negative sample division reference time, for example, the probability of conversion in the observation day, and the reference object content conversion degree and the possibility that the reference object is converted in the positive and negative sample division reference time form a positive correlation. If the training sample corresponding to the reference object is a negative sample, the reference object content conversion degree refers to the possibility that the reference object does not convert when the positive and negative samples are divided into reference time. The content conversion degree of the reference object is in positive correlation with the possibility that the reference object does not convert when the positive and negative samples are divided into reference time.

Specifically, the server may determine a sample type corresponding to the sample extraction feature, and if the sample type corresponding to the sample extraction feature is a reference sample type, input the sample extraction feature into the reference object identification network for identification.

In some embodiments, the reference object identification network may include a conversion likelihood identification network and a likelihood attenuation factor identification network. The input of the conversion possibility degree identification network is sample extraction characteristics, the output is total conversion possibility degree, the input of the possibility degree attenuation factor identification network is sample extraction characteristics, the output is possibility degree attenuation factors, the server can input the sample extraction characteristics into the conversion possibility degree identification network to identify the conversion possibility degree, the total conversion possibility degree of the reference object is obtained, the sample extraction characteristics are input into the possibility degree attenuation factor identification network to identify the attenuation factors, the possibility degree attenuation factors corresponding to the reference object are obtained, and calculation is carried out based on the possibility degree attenuation factors and the total conversion possibility degree of the reference object, and the content conversion degree of the reference object corresponding to the reference object is obtained.

The second conversion possibility degree identification network mentioned below refers to a conversion possibility degree identification network in the reference object identification network, and the second possibility degree attenuation factor identification network refers to a possibility degree attenuation factor identification network in the reference object identification network.

In some embodiments, the server processes the sample extraction features by using the network parameters of the second conversion possibility identification network to obtain processed sample extraction features, performs normalization processing on the processed sample extraction features, and uses the normalization result as the total conversion possibility of the reference object. For example, the server may calculate the total conversion likelihood using equation (1). Wherein, X _i The method includes the steps of representing sample extraction features corresponding to an object i, C being used for representing whether the object is converted or not finally, C =1 representing that a user is converted finally, C =0 representing that the user is not converted finally, X representing a user feature set and including user features X corresponding to a plurality of users respectively _i ，w _c The network parameters representing the conversion likelihood identify the network. P (X) _i ) Indicating the overall conversion probability.

In some embodiments, the training sample corresponding to the reference object is a positive sample, and the server may obtain the conversion occurrence duration corresponding to the reference object, and calculate the sub-conversion possibility according to the conversion occurrence duration and the possibility attenuation factor. Wherein when the conversion probability of the object gradually decays with the lapse of time and the decay conforms to the exponential distribution, the sub-conversion probability may be calculated using equation (2),

P(D＝d|X＝X _i ,C＝1)＝λ(X _i )·exp(-λ(X _i )d) (2)。

wherein, λ (X) _i ) Representing the likelihood attenuation factor and d the length of time that the conversion occurred. P (D = D | X = X) _i C = 1) indicates the subconvertibility and also indicates the probability of the conversion occurring on day d for the subject who will eventually convert. The server may perform a multiplication operation on the attenuation coefficient of the likelihood degree and the total conversion likelihood degree of the reference object, and use the result of the multiplication operation as the content conversion degree of the reference object, for example, the content conversion degree of the reference object may be calculated by using formula (3),

P(Y＝1,D＝d _i |X＝X _i ,E＝e _i )＝P(X _i )·λ(X _i )·exp(-λ(X _i )d _i ) (3)

wherein, P (Y =1, D = d) _i |X＝X _i ,E＝e _i ) The degree of conversion of the content of the reference object, i.e. the degree of conversion of the content of the reference object of the positive sample, can be expressed as P (X) _i )·λ(X _i )·exp(-λ(X _i )d _i )，d _i The length of time that the transformation for subject i occurs. D represents a conversion occurrence duration, for example, a time interval between release to the user and final conversion, Y represents whether the user has converted by the time of observation (it can be understood that the positive and negative samples divide the reference time as described above), Y =1 represents that the user has converted by the time of observation, Y =0 represents that the user has not converted by the time of observation, and E represents an observation duration, for example, a time interval between release and current waiting. For example, if a user puts a red dot 5 month and 1 day, and finally converts 5 month and 5 days, and the user has not converted yet when viewing the user at 5 month and 3 days, the values corresponding to the symbols are: y =0, c =1, d =5, e =3.

In some embodiments, the training sample corresponding to the reference object is a negative sample, the server may obtain an observation duration, where the observation duration refers to a time length corresponding to the observation time period, the server may calculate a likelihood adjustment value of the reference object according to the observation duration, a total conversion likelihood of the reference object, and a likelihood attenuation factor, and obtain the likelihood adjustment value based on the likelihood adjustment value of the reference object and the total conversion likelihoodThe reference object content conversion degree is calculated, for example, the server may calculate the reference object content conversion degree by using formula (4). Wherein P (Y =0 calory x = x _i ,E＝e _i ) Indicating the conversion degree of the content of the reference object, i.e. the conversion degree of the content of the reference object corresponding to the negative sample is 1-P (X) _i )+P(X _i )exp(-λ(X _i )e _i )，P(X _i )exp(-λ(X _i )e _i ) Indicating a likelihood adjustment value.

S212, a pushing identification loss value is obtained based on the pushing object content conversion degree corresponding to the content pushing object, and a reference identification loss value is obtained based on the reference object content conversion degree.

The push identification loss values are calculated according to the push object content conversion degrees of the content push objects, and may be calculated according to the push object content conversion degrees of one content push object, that is, one push object content conversion degree may correspond to one push identification loss value, and the push identification loss values may also be calculated according to the push object content conversion degrees of a plurality of content push objects, where a plurality means at least two. The push identification loss value and the push object content conversion degree are in a negative correlation relation.

The reference recognition loss value is calculated according to the reference object content conversion degree of the reference object, and may be calculated according to the reference object content conversion degree of one reference object, that is, one reference object content conversion degree may correspond to one reference recognition loss value, and the reference recognition loss value may also be calculated according to the reference object content conversion degrees of a plurality of reference objects. The reference identification loss value is in a negative correlation with the degree of conversion of the reference object content.

The negative correlation relationship refers to: under the condition that other conditions are not changed, the changing directions of the two variables are opposite, and when one variable is changed from large to small, the other variable is changed from small to large. It is understood that the negative correlation herein means that the direction of change is reversed, but it is not required that when one variable changes at all, the other variable must also change.

Specifically, the server may perform a statistical operation on the push object content conversion degrees of the content push objects to obtain the push statistical conversion degrees, for example, the server may perform an addition operation on the push object content conversion degrees and use the result of the addition as the push statistical conversion degrees, for example, the server may perform a product operation on the push object content conversion degrees and use the result of the product operation as the push statistical conversion degrees, or the server may perform a product operation on the push object content conversion degrees and perform a logarithm operation on the result of the product operation and use the result of the logarithm operation as the push statistical conversion degrees. The server can obtain a push identification loss value based on the push statistical conversion degree, and the push identification loss value and the push statistical conversion degree form a negative correlation relation.

In some embodiments, the server may perform a statistical operation on the reference object content conversion degrees of the respective reference objects to obtain the reference statistical conversion degrees, for example, the server may perform an addition operation on the reference object content conversion degrees and use a result of the addition as the reference statistical conversion degrees, for example, the server may perform a product operation on the reference object content conversion degrees and use a result of the product operation as the reference statistical conversion degrees, or the server may perform a product operation on the reference object content conversion degrees and perform a logarithm operation on a result of the product operation and use a result of the logarithm operation as the reference statistical conversion degrees. The server may obtain a push recognition loss value based on the reference statistical conversion degree, the push recognition loss value and the reference statistical conversion degree being in a negative correlation relationship. For example, the server may calculate a push identification loss value using equation (5), where L _DFM Indicating a push identification loss value.

L _DFM ＝-∑ _y＝1 logP(X _i )+logλ(X _i )-λ(X _i )d _i -∑ _y＝0 log(1-P(X _i )+P(X _i )exp(-λ(X _i )e _i ))(5)

S214, performing parameter adjustment on the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and obtaining a trained object recognition model based on the adjusted shared feature extraction network.

The shared feature extraction network may correspond to a plurality of parameters, the almanac of the shared feature extraction network refers to variable parameters inside the shared feature extraction network, and for the neural network, the parameters may also be referred to as bible network weights (weights). The trained object recognition model may be trained one or more times.

Specifically, the server may adjust the parameter of the shared feature extraction network toward a direction in which the push recognition loss value becomes smaller, and adjust the parameter of the shared feature extraction network toward a direction in which the reference recognition loss value becomes smaller, and may perform iterative training for multiple times, and stop the training when a convergence condition is satisfied, to obtain a trained object recognition model. The convergence condition may include an exponential one of the reference recognition loss value being less than the reference loss value threshold, the pushed recognition loss value being less than the pushed recognition loss value threshold, or a sum of the reference recognition loss value and the pushed recognition loss value being less than the total loss value threshold. The convergence condition may further include that the change of the parameter is less than a parameter change threshold. The reference loss value threshold, the push loss value threshold, the total loss value threshold, and the parameter change threshold may be preset or set as needed.

In some embodiments, the server may perform parameter adjustment on the shared feature extraction network sequentially using the pushed recognition loss value and the reference recognition loss value, for example, the server may perform parameter adjustment on the shared feature extraction network first using the pushed recognition loss value to obtain the shared feature extraction network after the pushed recognition loss value is adjusted, then perform parameter adjustment on the shared feature extraction network after the pushed recognition loss value is adjusted using the reference recognition loss value to obtain the shared feature extraction network after the reference recognition loss value is adjusted, or may perform parameter adjustment on the shared feature extraction network first using the reference recognition loss value to obtain the shared feature extraction network after the reference recognition loss value is adjusted using the pushed recognition loss value, then perform parameter adjustment on the shared feature extraction network after the reference recognition loss value is adjusted using the pushed recognition loss value to obtain the shared feature extraction network after the pushed recognition loss value is adjusted, and obtain the trained object recognition model based on the adjusted shared feature extraction network.

In some embodiments, there are multiple push identification loss values, for example, each push object content conversion degree may be calculated to obtain one push identification loss value, there are multiple reference identification loss values, for example, each reference object content conversion degree may be calculated to obtain one reference identification loss value, and the server may sequentially perform parameter adjustment on the shared feature extraction network by using the multiple push identification loss values and the multiple reference identification loss values, and obtain a trained object identification model based on the adjusted shared feature extraction network.

In some embodiments, the server may perform statistical calculation on the reference recognition loss value and the pushed recognition loss value to obtain a statistical loss value, perform parameter adjustment on the shared feature extraction network by using the statistical loss value, and obtain a trained object recognition model based on the adjusted shared feature extraction network.

In the object identification method, because the sample extraction features corresponding to the training samples of the content push object are extracted by the shared feature extraction network, and the sample extraction features corresponding to the training samples of the reference object are also extracted by the shared feature extraction network, the training samples of the content push object and the training samples of the reference object share one shared feature extraction network, so that under the condition of unbalanced sample number, the shared feature extraction network can be trained by two different samples, namely the training samples of the content push object and the training samples of the reference object, so that the shared feature extraction network learns the capability of distinguishing the two samples, the effect of model training is improved, the accuracy of the trained object identification model is improved, and the accuracy of object identification is improved.

In some embodiments, determining the target sample type corresponding to the sample extraction feature comprises: acquiring a sample mark value of a training sample corresponding to the sample extraction characteristic; and determining the type of the target sample corresponding to the sample extraction features based on the sample mark value.

Each training sample can correspond to a sample label value, the sample label values are used for distinguishing sample types, the sample label values are different, and the corresponding sample types are different. The sample flag value corresponding to the push sample type may be a value greater than or equal to 1, and may be 1, for example. The sample flag value corresponding to the reference sample type may be a numerical value less than or equal to 0, and may be 0, for example.

Specifically, the server may input the training sample and a sample label value corresponding to the training sample into the object recognition model, determine a sample type of the training sample according to the sample label value of the training sample when the sample extraction feature corresponding to the training sample is obtained, input the sample extraction feature into a push object recognition network in the object recognition model for recognition if the sample type is the push sample type, and input the sample extraction feature into a reference object recognition network in the object recognition model for recognition if the sample type is the reference sample type.

In this embodiment, a sample flag value of a training sample corresponding to a sample extraction feature is obtained, a target sample type corresponding to the sample extraction feature is determined based on the sample flag value, and a push training sample and a reference training sample in a training sample set can be distinguished, so that although the push training sample and the reference training sample share a shared feature extraction network, an object recognition model can still distinguish the push training sample and the reference training sample.

In some embodiments, the content conversion degree of the push object is used to indicate a conversion possibility degree of the content push object for the push content, and obtaining the push identification loss value based on the content conversion degree of the push object corresponding to the content push object includes: performing product calculation on the sample marking value and the content conversion degree of the push object to obtain an object identification loss value corresponding to the content push object; and counting the object identification loss values of the plurality of content push objects to obtain the push identification loss values.

The conversion possibility degree refers to the possibility degree of conversion of the object, such as the possibility degree of the object purchasing a product in the pushed advertisement or the possibility degree of the object clicking the pushed advertisement. When the training sample of the content push object is a positive sample, the push object content conversion degree represents the conversion possibility of the content push object for the push content, that is, the push object content conversion degree represents the possibility of the content push object converting at the positive and negative sample division reference time. The greater the likelihood of a conversion, the greater the probability of the user developing a conversion.

Specifically, the server may multiply the sample flag value by the push object content conversion degree to obtain a multiplied push object content conversion degree, obtain an object identification loss value of the content push object based on the multiplied push object content conversion degree, where the object identification loss value and the multiplied push object content conversion degree are in a negative correlation relationship, and the server may sum the object identification loss values of the plurality of content push objects and use a result of the sum as the push identification loss value.

In some embodiments, the server may determine, based on the content conversion degree of the push object, a first initial object loss value corresponding to the content push object, where the first initial object loss value and the content conversion degree of the push object are in a negative correlation relationship, perform a weighted calculation on each first initial identification loss value by using the sample flag value, and use a result of the weighted calculation as the push identification loss value. For example, the push identification loss value may be calculated using equation (6), where L _T Indicating a push recognition loss value, S ^T Representing a content push object set, S representing a training sample set, k ⁱ Sample label value, k, representing training sample corresponding to object i ⁱ =1 denotes sample type push sample type, k ⁱ =0 indicates that the sample type is the reference sample type. K includes a number of values of 0 and 1,

a first initial object loss value, S, representing object i _i Representing the training samples corresponding to the object i in the training sample set S.

In some embodiments, the server may determine a second initial object loss value corresponding to the reference object based on the content conversion degree of the reference object, where the second initial object loss value is in a negative correlation with the content conversion degree of the reference object, and the server may calculate the reference identification loss value by using the sample tag value and the second initial object loss value. For example, the server may calculate a reference recognition loss value using equation (7) where L _C Representing a reference recognition loss value, S ^C Representing a set of reference objects, S comprising S ^T And S ^C 。

In this embodiment, the product of the sample tag value and the content conversion degree of the push object is calculated to obtain an object identification loss value corresponding to the content push object, and the object identification loss values of the plurality of content push objects are counted to obtain a push identification loss value, so that the efficiency of calculating the loss value is improved.

In some embodiments, performing parameter adjustment on the shared feature extraction network based on the pushed recognition loss value and the reference recognition loss value, and obtaining the trained object recognition model based on the adjusted shared feature extraction network comprises: performing parameter adjustment on the shared feature extraction network based on the pushing identification loss value to obtain the shared feature extraction network after the pushing identification loss value is adjusted; performing parameter adjustment on the shared feature extraction network subjected to pushing identification loss value adjustment based on the reference identification loss value to obtain a reference identification loss value and the shared feature extraction network subjected to pushing identification loss value adjustment; and extracting a network based on the reference recognition loss value and the shared feature adjusted by the push recognition loss value to obtain a trained object recognition model.

Specifically, the server may input the push training sample into the object recognition model to obtain a push object content conversion degree of the content push object, obtain a push recognition loss value based on the push object content conversion degree of the content push object, perform parameter adjustment on the shared feature extraction network based on the push recognition loss value to obtain a shared feature extraction network adjusted based on the push recognition loss value, input the reference training sample into the object recognition model to obtain a reference object content conversion degree of the reference object, obtain a reference recognition loss value based on the reference object content conversion degree of the reference object, perform parameter adjustment on the shared feature extraction network adjusted based on the push recognition loss value based on the reference recognition loss value to obtain a shared feature extraction network adjusted based on the reference recognition loss value and the push recognition loss value, and obtain a trained object recognition model based on the shared feature extraction network adjusted based on the reference recognition loss value and the push recognition loss value. The shared feature extraction network can perform parameter adjustment through the push identification loss value for multiple times or through the reference identification loss value for multiple times, and the sequence of performing parameter adjustment by using the push identification loss value and performing parameter adjustment by using the reference identification loss value can be set as required without excessive limitation.

In this embodiment, the shared feature extraction network is adjusted by using the push recognition loss value and the parameters of the shared feature extraction network are adjusted by using the reference recognition loss value, so that the trained shared feature extraction network learns the features of the content push object and the features corresponding to the reference, thereby improving the training accuracy.

In some embodiments, performing parameter adjustment on the shared feature extraction network based on the pushed recognition loss value to obtain the shared feature extraction network after the pushed recognition loss value is adjusted includes: determining a push parameter adjustment related value corresponding to the push object identification network based on the push identification loss value, and performing parameter adjustment on the push object identification network based on the push parameter adjustment related value to obtain a push object identification network after parameter adjustment; determining a first characteristic parameter adjustment correlation value corresponding to the shared characteristic extraction network based on the push identification loss value and the first loss value adjustment direction; and performing parameter adjustment on the shared feature extraction network based on the first feature parameter adjustment related value to obtain the shared feature extraction network with the adjusted push identification loss value.

The push parameter adjustment related value is a value used for adjusting parameters of the push object identification network, the push identification loss value corresponds to a push identification loss function, and the push identification loss value is a specific value of the push identification loss function. The first loss value adjustment direction may be a direction in which the gradient decreases. The first feature parameter adjustment related value refers to a numerical value used for adjusting a parameter of the shared feature extraction network.

Specifically, the server may adjust parameters in the shared feature extraction network toward a direction in which the pushed recognition loss value becomes smaller, and adjust parameters in the pushed object recognition network toward a direction in which the pushed recognition loss value becomes smaller, to obtain an adjusted shared feature extraction network and an adjusted pushed object recognition network, and obtain a trained object recognition model based on the adjusted shared feature extraction network and the adjusted pushed object recognition network.

In some embodiments, the server may perform a gradient operation on the parameter of the push object identification network by using a push identification loss function to obtain a first gradient, and determine a push parameter adjustment correlation value based on the first gradient, where adjusting the parameter of the push object identification network by using the push parameter adjustment correlation value may change the first gradient toward a descending direction.

In some embodiments, the server may perform a gradient operation on the parameter of the shared feature extraction network by using a push identification loss value to obtain a second gradient, where a first loss value adjustment direction is a direction in which the second gradient decreases, and the server may determine a first feature parameter adjustment correlation value based on the second gradient, where adjusting the parameter of the shared feature extraction network by using the first feature parameter adjustment correlation value may change the second gradient toward the decreasing direction. The server may adjust a parameter of the shared feature extraction network in a direction in which the second gradient is decreased, to obtain the shared feature extraction network with the adjusted push identification loss value.

In the embodiment, the push object identification network and the shared feature extraction network are adjusted by using the push identification loss value, and one loss value adjusts two networks, so that the efficiency of parameter adjustment is improved.

In some embodiments, performing parameter adjustment on the pushed recognition loss value-adjusted shared feature extraction network based on the reference recognition loss value to obtain the reference recognition loss value and the pushed recognition loss value-adjusted shared feature extraction network includes:

determining a reference parameter adjustment correlation value corresponding to the reference object identification network based on the reference identification loss value; performing parameter adjustment on the reference object identification network based on the reference parameter adjustment related value to obtain the reference object identification network after the parameter adjustment; determining a second characteristic parameter adjustment related value corresponding to the shared characteristic extraction network after the pushing identification loss value is adjusted based on the reference identification loss value and the second loss value adjustment direction; and performing parameter adjustment on the shared feature extraction network after the push identification loss value is adjusted based on the second feature parameter adjustment related value to obtain a reference identification loss value and the shared feature extraction network after the push identification loss value is adjusted.

The reference parameter adjustment correlation value is a numerical value used for adjusting a parameter of the reference object identification network, the reference identification loss value corresponds to a reference identification loss function, the reference identification loss value is a specific value of the reference identification loss function, and the second loss value adjustment direction may be a gradient descending direction. The second feature parameter adjustment related value refers to a numerical value used for adjusting a parameter of the shared feature extraction network.

Specifically, the server may adjust parameters in the shared feature extraction network toward a direction in which the reference recognition loss value becomes smaller, and adjust parameters in the reference object recognition network toward a direction in which the reference recognition loss value becomes smaller, resulting in an adjusted shared feature extraction network and an adjusted reference object recognition network, and obtaining a trained object recognition model based on the adjusted shared feature extraction network and the adjusted reference object recognition network.

In some embodiments, the server may perform a gradient operation on the parameter of the reference object identification network by using a reference identification loss function to obtain a third gradient, and determine a reference parameter adjustment correlation value based on the third gradient, wherein adjusting the parameter of the reference object identification network by using the reference parameter adjustment correlation value may change the third gradient toward a descending direction.

In some embodiments, after the shared feature extraction network is adjusted by using the pushed recognition loss value, the server may perform a gradient operation on the parameter of the shared feature extraction network by using a reference recognition loss value to obtain a fourth gradient, where a second loss value adjustment direction is a direction in which the fourth gradient decreases, and the server may determine a second feature parameter adjustment correlation value based on the fourth gradient, where adjusting the parameter of the shared feature extraction network by using the second feature parameter adjustment correlation value may change the fourth gradient toward the decreasing direction. The server may adjust a parameter of the shared feature extraction network in a direction in which the fourth gradient is decreased, to obtain a reference recognition loss value and push the shared feature extraction network after the recognition loss value is adjusted.

In the embodiment, the reference object identification network and the shared feature extraction network are adjusted by using the reference identification loss value, and one loss value adjusts two networks, so that the parameter adjustment efficiency is improved.

In some embodiments, the push object identification network comprises a translation likelihood identification network and a likelihood attenuation factor identification network; inputting the sample extraction characteristics of which the target sample type is the push sample type into a push object identification network in an object identification model for identification, wherein the identification for obtaining the content conversion degree of the push object corresponding to the content push object comprises the following steps: extracting characteristics of a sample with a target sample type as a push sample type, and inputting the characteristics into a conversion possibility degree identification network in a push object identification network for conversion possibility degree identification to obtain a total conversion possibility degree; extracting characteristics of a sample with a target sample type as a push sample type, and inputting the characteristics into a likelihood attenuation factor identification network in a push object identification network for attenuation factor identification to obtain a likelihood attenuation factor; a push object content conversion degree of the content push object is determined based on the total conversion likelihood and the likelihood attenuation factor.

Specifically, the server may input the sample extraction features into the attenuation factor recognition network, and process the sample extraction features using the network parameters of the attenuation factor recognition network to obtain the attenuation factor of the likelihood of the content push object, where the attenuation factor of the likelihood may be, for example, a

Wherein, w _d Network parameters of the network are identified for the likelihood attenuation factor.

In some embodiments, the training sample of the content push object is a positive sample, and the server may obtain a conversion occurrence duration, where the conversion occurrence duration refers to a time length between the content push time and the conversion occurrence time. Transformation occurrence time refers to the time at which transformation occurs in a subject. The server may calculate a sub-conversion possibility of the content push object according to the observation duration and the possibility attenuation factor, and calculate a push object content conversion degree of the content push object based on the sub-conversion possibility of the content push object and the total conversion possibility of the content push object, for example, the server may perform a product operation on the sub-conversion possibility and the total conversion possibility, and use a result of the product operation as the push object content conversion degree.

For example, if the server pushes content to the content push object on 1 month and 1 day, the positive/negative sample division reference time is 1 month and 10 days, if the content push object is converted on 1 month and 6 days, the content push object is a positive sample, and if the content push object has a subconversibility of 0.6 on 1 month and 10 days and a total conversion possibility of the content push object is 0.8, the push object content conversion degree of the content push object is 0.6 × 0.8=0.48.

In some embodiments, the training sample of the content push object is a negative sample, the server may obtain an observation duration, where the observation duration refers to a time length corresponding to the observation time period, the server may calculate a likelihood adjustment value of the content push object according to the observation duration, a total conversion likelihood of the content push object, and a likelihood attenuation factor, calculate a content conversion degree of the content push object based on the likelihood adjustment value of the content push object and the total conversion likelihood,

in some embodiments, the push object recognition network may be implemented using a Delayed Feedback Model (DFM). The reference object recognition network may also employ a delayed feedback model. As shown in fig. 3, a model structure diagram of a delayed feedback model is shown, which includes an embedding Layer, a polymerization Layer (conditioner), a conversion Layer (Cvr Layer), and a parameter Layer (Lambda Layer). As shown in fig. 4, the block diagram of the object recognition model in some embodiments includes a shared feature extraction network, a pushed object recognition network, and a reference object recognition network, where the shared feature extraction network includes an embedding layer, an aggregation layer, and a feature extraction layer, the pushed object recognition network includes a first transformation likelihood recognition network and a first likelihood attenuation factor recognition network, and the reference object recognition network includes a second transformation likelihood recognition network and a second likelihood attenuation factor recognition network.

In this embodiment, the pushing object content conversion degree is obtained by using the probability attenuation factor and the total conversion probability, and since the probability attenuation factor can extract the time-lapse attenuation of the probability of the object that the conversion occurs, the time-lapse attenuation of the probability of the conversion occurring in the calculated pushing object content conversion degree is also extracted, so that the accuracy of the pushing object content conversion degree is improved.

In some embodiments, determining a push object content conversion for the content push object based on the total conversion likelihood and the likelihood attenuation factor comprises: acquiring possible attenuation duration; determining a subconvertibility degree of the content pushing object based on the attenuation duration of the possibility degree and the attenuation factor of the possibility degree, wherein the subconvertibility degree is used for reflecting the possibility degree of conversion of the object corresponding to the time when the time interval between the content pushing time and the attenuation duration of the possibility degree is the attenuation duration of the possibility degree; and determining the content conversion degree of the push object of the content push object based on the sub-conversion possibility degree and the total conversion possibility degree.

Wherein the time-of-likelihood decay time period may be any one of an observation time period or a time period during which the conversion occurs. When the training sample is a positive sample, the possible degree attenuation time length is the conversion occurrence time length, and when the training sample is a negative sample, the possible degree attenuation time length is the observation time length.

Specifically, when the training sample is a positive sample, the server may obtain a conversion occurrence duration corresponding to the content push object, and use the conversion occurrence duration as a likelihood attenuation duration, and the server may determine a subconvertion likelihood of the content push object based on the conversion occurrence duration and the likelihood attenuation factor, where the subconvertion likelihood indicates a likelihood that the content push object is converted at a time when a time interval between the content push object and the content push time is the conversion occurrence duration, for example, if the content push time is 1 month and 1 day and the conversion occurrence duration is 4 days, the subconvertion likelihood may indicate a likelihood that the content push object is converted at 1 month and 5 days. The server may perform a product operation on the sub-conversion possibility and the total conversion possibility of the content push object, and use a result of the operation as the push object content conversion degree of the content push object. In some embodiments, when the training sample is a negative sample, the server may obtain the observation duration as a possible degree decay duration, and for an object whose reference time has not been converted by dividing the positive and negative samples, i.e., for a negative sample, there are two possibilities, one is that the object will not be converted at all, i.e., C =0, and one is that the object will be converted at all, i.e., C =1, but since the observation duration is less than the conversion occurrence duration of the object, the object is not found to be converted by the observation day. Wherein, the observation day refers to the reference time for dividing positive and negative samples. Therefore, when the training sample is a negative sample, the possibility of the content push object being converted on the observation day, that is, the push object content conversion degree of the content push object, can be represented by formula (4).

In this embodiment, the attenuation of the probability of occurrence of conversion over time is also improved in the calculated conversion degree of the content of the push object, so that the accuracy of the conversion degree of the content of the push object is improved.

In some embodiments, obtaining a training sample set corresponding to an object recognition model to be trained includes: acquiring training push content; dividing a training object set to obtain a content pushing object set and a reference object set; pushing the training push content to each content push object in a content push object set, and acquiring the content push time of the training push content, wherein a reference object in a reference object set shields the automatic push training push content; acquiring a backward operation behavior record of a content push object aiming at training push content; determining a content conversion result of a content push object in the content push objects based on the backward operation behavior record; for a reference object in the reference object set, acquiring a content conversion result of the reference object after the content push time; and taking the training sample corresponding to the object with the content conversion result of converted as a positive sample in the training sample set, and taking the training sample corresponding to the object with the content conversion result of unconverted as a negative sample in the training sample set.

Wherein the training push content is content pushed to a content push object. The training object set includes a plurality of objects, and each object may be a user on the same platform, for example, a user on the same shopping platform. The content push time is a time for pushing the training push content to the content push object. The backward operation behavior record for the training push content refers to a behavior of the content push object for the training push content after the content push time and before the positive and negative sample division reference time (which may include the positive and negative sample division reference time), and may include a behavior of clicking on the training push content or purchasing a commodity related to the training push content, for example. The content conversion result may be either converted or unconverted.

Specifically, the server may divide the training object set to obtain a content push object set and a reference object set, where the number of objects included in the content push object set may be the same as or different from the number of objects included in the reference object set, for example, the number of objects included in the content push object set may be less than the number of objects included in the reference object set. The server actively pushes the training push content to the content push object, and does not actively push the training push content to the reference object. The server may determine whether the conversion behavior is included in the backward operation behavior record, determine that the content conversion result of the content push object is converted when the conversion behavior is included, and determine that the content conversion result of the content push object is unconverted when the conversion behavior is not included in the backward operation behavior record. The conversion behavior is used to determine whether conversion occurs, for example, when a user has a purchase behavior, it is determined that the user has conversion, and the conversion behavior may be a purchase.

In some embodiments, the server may obtain an object behavior of the reference object after the content push time, determine that the content conversion result of the reference object is converted when the object behavior includes a conversion behavior, and determine that the content conversion result of the reference object is unconverted when the object behavior does not include a conversion behavior, for example, when the training push content is an advertisement of a computer a, if a user purchases the computer a, it indicates that the user has converted, and when the user has a behavior of purchasing the computer a, it determines that the content conversion result of the user is converted.

In some embodiments, the server may use, as the positive samples in the training sample set, the training samples whose content conversion results are corresponding to the converted objects before the positive and negative sample division reference time, and use, as the negative samples in the training sample set, the training samples whose content conversion results are corresponding to the unconverted objects before the positive sample division reference time. The positive and negative samples divide the reference time after between content pushes.

In this embodiment, the positive sample and the negative sample are determined by an operation behavior after the content pushing time, so that the efficiency and the accuracy of obtaining the training sample are improved.

In some embodiments, as shown in fig. 5, an object recognition method is provided, in which the trained object recognition model generated in the above embodiments is applied to perform object recognition, which is described by taking the application of the method to the server 104 in fig. 1 as an example, and includes the following steps:

s502, acquiring a candidate object set corresponding to the target push content; the candidate set comprises a plurality of candidate objects.

The target push content may be any content, may be preset or determined according to needs, and may be a coupon, for example. The target push content may be the same as the training push content or may be different from the training push content. The set of candidate objects may be pre-stored in the server. The content push request may be automatically triggered periodically by the server or obtained from another device.

Specifically, the server may obtain the candidate object set in response to a content push request for the target push content, where the content push request is used to request that the target push content be pushed to the object. The server can determine a push object from the candidate object set and push the target push content to the push object, wherein the push object comprises a marketing sensitive object, namely an object which does not generate conversion when content push is not performed and an object which generates conversion when content push is performed.

S504, inputting the object sample of the candidate object into the shared feature extraction network of the trained object recognition model for feature extraction, and obtaining the object extraction feature corresponding to the candidate object.

The object sample refers to a sample corresponding to the candidate object, and the server may generate the object sample according to the object attribute information of the candidate object. The object extraction features are features obtained by performing feature extraction on an object sample by using a shared feature extraction network.

S506, inputting the object extraction features into a pushed object recognition network in an object recognition model for recognition, and obtaining a first content conversion possibility corresponding to the candidate object based on the total conversion possibility obtained through recognition.

The first content conversion possibility degree represents the possibility degree of conversion of the candidate object within the target possibility degree attenuation duration, and the first content conversion possibility degree is calculated according to a result obtained by identifying the object extraction features by the pushed object identification network. The target time duration of decay may be set as desired, and may be, for example, 10 days.

Specifically, the pushed object identification network includes a first conversion possibility identification network and a first possibility attenuation factor identification network, the server may process the object extraction feature by using the first conversion possibility identification network to obtain a first total conversion possibility of the candidate object, and process the object extraction feature by using the first possibility attenuation factor identification network to obtain a first possibility attenuation factor of the candidate object, the server may obtain a target possibility attenuation duration, and calculate a first content conversion possibility corresponding to the candidate object based on the target possibility attenuation duration, the first possibility attenuation factor, and the first total conversion possibility. For example, the server may calculate the first content conversion possibility corresponding to the candidate object by using formula (8).

Where d represents the target likelihood attenuation duration, which may be, for example, 5 days, k is the time in the target likelihood attenuation duration, e.g., k =1 represents day 1 of 5 days, pcvr represents the first total conversion likelihood, λ (X) _i ) Representing a first degree of probability attenuation factor, X _i Representing object extraction features. pCVR _ d represents the first content conversion possibility.

And S508, inputting the object extraction features into a reference object recognition network in the object recognition model for recognition, and obtaining a second content conversion possibility corresponding to the candidate object based on the conversion possibility obtained through recognition.

The second content conversion possibility degree represents the possibility degree of conversion of the candidate object within the target possibility degree attenuation duration, and is calculated according to a result obtained by identifying the object extraction features by the reference object identification network.

Specifically, the reference object identification network includes a second conversion likelihood identification network and a second likelihood attenuation factor identification network, the server may process the object extraction features by using the second conversion likelihood identification network to obtain a second total conversion likelihood of the candidate object, process the object extraction features by using the second likelihood attenuation factor identification network to obtain a second likelihood attenuation factor of the candidate object, the server may obtain the target likelihood attenuation duration, and calculate a second content conversion likelihood corresponding to the candidate object based on the target likelihood attenuation duration, the second likelihood attenuation factor, and the second total conversion likelihood. For example, the server may calculate the second content conversion possibility corresponding to the candidate object by using formula (8).

S510, obtaining the difference of the possibility degree between the first content conversion possibility degree and the second content conversion possibility degree corresponding to the candidate object.

Wherein, the difference of the possibility degree refers to the difference between the first content conversion possibility degree and the second content conversion possibility degree. The server may take the result of subtracting the second content conversion possibility from the first content conversion possibility as the possibility difference.

In some embodiments, a structure diagram of the trained object recognition model is shown in fig. 4, as shown in fig. 4, the server inputs the object extraction features into the first conversion likelihood recognition network to obtain a first total conversion likelihood, inputs the object extraction features into the first likelihood attenuation factor recognition network to obtain a first likelihood attenuation factor, calculates a first content conversion likelihood based on the first total conversion likelihood, the first likelihood attenuation factor and the target likelihood attenuation duration, and uses G in the diagram ^T (Y _i |X _i ) Indicating a first content conversion likelihood. Similarly, the server inputs the object extraction features into a second conversion possibility degree identification network to obtain a second total conversion possibility degree, inputs the object extraction features into a second possibility degree attenuation factor identification network to obtain a second possibility degree attenuation factor, calculates a second content conversion possibility degree based on the second total conversion possibility degree, the second possibility degree attenuation factor and the target possibility degree attenuation duration, and uses G in the graph ^C (Y _i |X _i ) Indicating a second content conversion probability. The server calculates the difference between the first content conversion possibility and the second content conversion possibility to obtain the difference of the possibility, wherein tau (X) is used in the figure _i )＝G ^T (Y _i |X _i )-G ^C (Y _i |X _i ) Denotes the difference in probability, τ (X) _i ) Represents a difference in likelihood, X _i Representing object extraction features.

S512, screening candidate objects meeting the difference condition from the candidate object set based on the difference of the corresponding possibility degrees of the candidate objects, and taking the candidate objects as push objects corresponding to target push contents; the difference condition includes at least one of a ranking of the energy difference before a preset ranking or a likelihood difference greater than a difference threshold.

The ranking refers to the ranking of the candidate objects in the candidate object sequence, the candidate object sequence is a sequence obtained by arranging the candidate objects in the candidate object set according to the sequence of the small degree difference, and the greater the degree difference is, the more the ranking of the candidate objects in the candidate object sequence is. The preset ordering may be set as desired or preset, and may be 10, for example. The difference threshold may be preset or set as desired.

Specifically, the server may compare the difference in the likelihood of the candidate object with a difference threshold, and when the difference in the likelihood is greater than the difference threshold, take the candidate object as a push object. Or, the server may select each candidate object ranked before the preset ranking from the candidate object sequence as the push object. The server can push target push content to the terminal corresponding to the push object.

The object identification method includes the steps of obtaining a candidate object set corresponding to target pushed content, wherein the candidate object set comprises a plurality of candidate objects, inputting an object sample of the candidate objects into a shared feature extraction network of a trained object identification model for feature extraction to obtain object extraction features corresponding to the candidate objects, inputting the object extraction features into a pushed object identification network of the object identification model for identification, obtaining a first content conversion possibility corresponding to the candidate objects based on the conversion possibility obtained through identification, inputting the object extraction features into a reference object identification network of the object identification model for identification, obtaining a second content conversion possibility corresponding to the candidate objects based on the conversion possibility obtained through identification, obtaining a possibility difference between the first content conversion possibility and the second content conversion possibility corresponding to the candidate objects, screening candidate objects meeting a difference condition from the candidate object set based on the possibility difference corresponding to the candidate objects to serve as the pushed objects corresponding to the target pushed content, and the difference condition comprises that the ranking of the possibility difference is before the preset ranking or the possibility difference is larger than a threshold value, so that the accuracy of the corresponding target pushed objects is improved by using the trained object identification model.

In some embodiments, the step of obtaining a trained object recognition model comprises: acquiring a training sample set corresponding to an object recognition model to be trained, wherein the training sample set comprises training samples corresponding to content push objects and training samples corresponding to reference objects; respectively inputting training samples in the training sample set into a shared feature extraction network in an object recognition model for feature extraction to obtain sample extraction features corresponding to the training samples; determining a target sample type corresponding to the sample extraction features; inputting the sample extraction characteristics of which the target sample type is the push sample type into a push object identification network in an object identification model for identification, and identifying to obtain a push object content conversion degree corresponding to a content push object; inputting sample extraction characteristics of which the target sample type is a reference sample type into a reference object identification network in an object identification model for identification, and identifying to obtain a reference object content conversion degree corresponding to a reference object; obtaining a pushing identification loss value based on a pushing object content conversion degree corresponding to a content pushing object, and obtaining a reference identification loss value based on a reference object content conversion degree corresponding to a reference object; and performing parameter adjustment on the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and obtaining a trained object recognition model based on the adjusted shared feature extraction network.

In this embodiment, because the sample extraction features corresponding to the training samples of the content push object are extracted by the shared feature extraction network, and the sample extraction features corresponding to the training samples of the reference object are also extracted by the shared feature extraction network, the training samples of the content push object and the training samples of the reference object share one shared feature extraction network, so that the shared feature extraction network can still accurately identify the samples under the condition of unbalanced sample number, thereby improving the sample identification accuracy and the accuracy of the object identification model.

The application also provides an application scene, and the application scene applies the object identification method. Specifically, the application of the object identification method in the application scenario is as follows:

1. and determining the target advertisement to be pushed.

2. Acquiring a candidate object set; the candidate set includes a plurality of candidates.

3. The method comprises the steps of obtaining a trained object recognition model, wherein the object recognition model comprises a shared feature extraction network, a pushed object recognition network and a reference object recognition network, the pushed object recognition network comprises a first conversion possibility degree recognition network and a first possibility degree attenuation factor recognition network, and the reference object recognition network comprises a second conversion possibility degree recognition network and a second possibility degree attenuation factor recognition network.

4. And inputting the object sample corresponding to the candidate object into a shared feature extraction network for feature extraction to obtain the object extraction feature corresponding to the candidate object.

5. And inputting the object extraction features into a first conversion possibility identification network to obtain a first total conversion possibility corresponding to the candidate object, and inputting the object extraction features into a first possibility attenuation factor identification network to obtain a first possibility attenuation factor corresponding to the candidate object.

6. And inputting the object extraction features into a second conversion possibility degree identification network to obtain a second total conversion possibility degree corresponding to the candidate object, and inputting the object extraction features into a second possibility degree attenuation factor identification network to obtain a second possibility degree attenuation factor corresponding to the candidate object.

7. And obtaining the attenuation duration of the target possibility, and obtaining the first content conversion possibility corresponding to the candidate object based on the first total conversion possibility, the first possibility attenuation factor and the attenuation duration of the target possibility.

8. And obtaining a second content conversion possibility corresponding to the candidate object based on the second total conversion possibility, the second possibility attenuation factor and the target possibility attenuation duration.

9. And acquiring the difference of the possibility between the first content conversion possibility and the second content conversion possibility corresponding to the candidate object.

10. And screening candidate objects meeting the difference condition from the candidate object set based on the difference of the corresponding possibility degrees of the candidate objects to serve as push objects corresponding to the target advertisements.

Wherein the difference condition comprises at least one of the ranking of the energy difference before the preset ranking or the likelihood difference being greater than a difference threshold.

11. And pushing the target advertisement to a terminal corresponding to the pushing object.

In the traditional method, marketing sensitive users can be screened through a gain model (uplift model), the modeling method of the gain model is different from that of a traditional response model (response model), and the gain model is used for predicting the gain of the probability of conversion of the users under certain intervention (treatment). The gain model can screen marketing-sensitive users by calculating a gain value, wherein the gain value refers to a difference value between the probability that the user is converted under the condition of being intervened and the probability that the user is converted under the condition of not being intervened.

Method for modeling gain model, e.g. tau _i ＝Y _i (1)-Y _i (0) (9) wherein i represents one user in a user group consisting of N users, and Y _i (1) Indicating the probability of user i having intervened on user i to make a transition, e.g. the probability of user i purchasing an item after issuing a coupon to user i, Y _i (0) The probability of the user conversion without the intervention of the user i is represented, such as the probability of the user i purchasing goods without issuing coupons to the user i. Tau is _i Is the difference between the probability of a transformation before and after an intervention, i.e. the difference between the probability of a transformation occurring for the user with an intervention and the probability of a transformation occurring for the user without an interventionValue, i.e. τ _i The gain value, which may also be referred to as an increment or a causal effect (cause effect) of the user i, represents an increase in the probability of a transition for the case with intervention compared to the case without intervention.

Since it is not possible for user i to be in the intervening and interfered states at the same time, for example, one can send a coupon to user i, or not send a coupon to user i, or send a coupon to user i and not send a coupon to user i, a conditional average causal effect (CATE) can be used to characterize the diversity of a population, i.e. a population's average causal effect is used instead of a human causal effect, and the formula of the conditional average causal effect is, for example, CATE:

τ(X _i )＝E[Y _i (1)|X _i ]-E[Y _i (0)|X _i ] (10)。

since it is a mutually exclusive event for the same user whether to intervene in the same event, the observed Y can be compared _i Rewritten as a formula

Y _i ^obs ＝W _i Y _i (1)+(1-W _i )Y _i (0)(11)，

When the user is characteristic X _i And the intervention condition W _i Independent of each other (Conditional Independence Assumption, CIA), the observed difference in causal effect can be expressed as equation (12).

τ(X _i )＝E[Y _i ^obs |X _i ＝x,W＝1]-E[Y _i ^obs |X _i ＝x,W＝0] (12)

In some embodiments, the modeling method of the gain Model may include at least One of a Two Model Approach (Two Model Approach), a single Model Approach (One Model Approach), or a Transformation Approach (Transformation Approach). The two-model approach includes an intervention group model and a control group model. As shown in FIG. 6, a block diagram of a process for training two models is shown. During training, the intervention group model and the control group model use independent samples, the intervention group model is trained by using the intervention group sample, and the control group sample is usedThe control group model was trained. The intervention group samples are samples corresponding to users who have been subjected to an intervention, and the control group samples are samples corresponding to users who have not been subjected to the intervention. As shown in fig. 7, a flow chart of gain value prediction using the trained two models is shown. During prediction, an intervention group model is used for predicting the conversion score of a prediction group sample to obtain an intervention conversion score, a control group model is used for predicting the conversion score of the prediction group sample to obtain a control conversion score, the difference value of the intervention conversion score and the control conversion score is calculated, and the calculated difference value is used as the conversion gain score corresponding to the prediction group sample. The intervention group model may be G ^T The control group model can be represented by G ^C The results obtained by the intervention group model may be represented by G ^T (Y _i |X _i ) The results obtained for the control group model can be expressed as G ^C (Y _i |X _i ) And (4) showing. In the two model approach, the conversion gain fraction (i.e., gain value) may be expressed as

τ(X _i )＝G ^T (Y _i |X _i )-G ^C (Y _i |X _i ) (13)。

The control group model and the intervention group model may be collectively referred to as a base model G, and the base model G may be implemented by using a classification or regression model, for example, a multilayer neural network based on cross entropy loss may be used to predict a predicted value such as a conversion rate of a user as a ranking basis. The intervention group may also be referred to as the experimental group.

In the two model methods, independent samples are used during the training of the intervention group model and the control group model, so that scoring errors are accumulated, and a single model method is provided, wherein whether intervention is performed or not is used as a characteristic in the training process. Specifically, in the single model method, the original user characteristics X are measured in the sample dimension _i The expansion is carried out, and characteristics T related to the intervention are introduced into the modeling, such as T =1 as the intervention group characteristics and T =0 as the control group characteristics. The model structure of the single model is shown in FIG. 8, and the output result of the single model can be G (Y) _i |X _i T) represents that, in a single model, the conversion gain fraction can be expressed as

τ(X _i )＝G(Y _i |X _i ,T＝1)-G(Y _i |X _i ,T＝0)(14)。

The category conversion method belongs to a single model method, can only aim at a binary intervention scene (if the intervention is not applied), and introduces an intermediate variable Z to optimize tau (X) in the category conversion method _i ). The model structure of the class conversion method is shown in fig. 9. The intermediate variable Z can be expressed by equation (15). I.e. Z =1 if the user is intervened and converted, Z =1 if the user is not intervened and converted, and Z =0 otherwise. The output result of the single model in the class conversion method may be G (Z) _i |X _i ) And (4) showing.

In the case where the number of samples in the intervention group and the control group is the same, i.e., the probability that the user is from the intervention group is equal to the probability from the control group, the conversion gain score can be expressed as formula (16), that is, τ (X) is learned _i ) Is equivalent to learning P [ Z ] _i ＝1|X _i ]The original label Y (Y in the formula) can be converted into Z by using the formula (15), the Z is optimized by using the model G, and the prediction is carried out according to the estimated P [ Z ] Z _i ＝1|X _i ]And (5) sorting.

In the class conversion method, the number of samples of the intervention group and the control group needs to be completely equal, which is difficult to satisfy in reality, and the two sample numbers are generally changed to be consistent by resampling, or the user tendency may be introduced to change the sample numbers of the two groups to be consistent. The class conversion method is only suitable for the binary condition of whether to intervene, and cannot be used if the intervention has multiple types (such as red packets with different money amounts).

Therefore, in the conventional modeling method of the gain model, samples of the intervention group and the comparison group need to obey condition independent assumption, and in order to satisfy the condition, it is practically necessary to randomize the a/B experiment on line to set the intervention group and the comparison group so that feature distributions of the two groups are consistent, for example, in the a/B experiment, by randomly extracting experimenters, the experimenters are divided into 2 groups, each group includes 25 persons, the group a uses the keyboard layout a, and the group B uses the keyboard layout B. They were asked to type a standard 20 word text message in 30 seconds and then the number of miswords was recorded. However, for the financial scene, due to low user conversion rate and low user behavior frequency, it is difficult to reserve a large unbiased comparison group, that is, unbiased comparison samples are scarce in the financial scene, the user decision cost is high, the conversion is difficult, and the proportion of positive and negative samples is very different. Therefore, the number of samples in the control group is far smaller than that of the samples in the intervention group, and the positive samples completely depend on the natural transformation of the user because the samples in the control group are not interfered, so that the proportion of the positive samples and the negative samples in the control group is more different, and the traditional modeling method of the gain model is not suitable for the financial scene. The reason is as follows: for the two-model method, the insufficient training of the comparison group model can be caused by the shortage of the comparison group sample, and the error is increased; for the single model method, the influence of the intervention characteristic T on the Loss value (Loss) of the model is small due to the insufficient samples of the control group, the intervention characteristic T in the model hardly plays an expected role, and the score difference values output by the model when T =1 and T =0 are predicted are small or even not different; for the class conversion method, since the class conversion method needs the number of samples in the intervention group to be equal to the number of samples in the control group, the amount of samples in the intervention group and the amount of samples in the control group are generally made to be consistent by resampling, or the amount of samples in the intervention group and the amount of samples in the control group are made to be consistent by introducing user tendency, however, the resampling method may cause the samples in the intervention group to become as rare as the samples in the control group, the introducing user tendency and the like may cause the prediction score to be mainly contributed by the intervention group, and the class conversion method may only be used in a scenario where the intervention factor is binary (e.g., whether the red packet is generated or not).

In the object identification method provided by the present application, the push training samples (i.e., the intervention group samples) and the reference training samples (i.e., the comparison group samples) share a shared feature extraction network, that is, the same shared feature extraction network is obtained through training of the intervention group samples and the comparison group samples, so that even if the number of the intervention group samples is inconsistent with the number of the comparison group samples, for example, in the case that the number of the intervention group samples is greater than the number of the comparison group samples, a network capable of locally intervening the group samples and the comparison group samples can still be obtained through training, and thus the object identification method provided by the present application can be applied to a scene with unbalanced samples, for example, a financial scene, for example, a scene such as advertisement push or coupon issuance in the financial scene.

In the conventional modeling method of the gain model, the base model G is usually a classification model based on cross entropy loss or a regression model based on square root loss, and does not consider the delay of user transformation. However, in a financial scenario, the decision of the user is greatly influenced by external factors such as market conditions, the decision making cost of the user for subscription is high, and the conversion is delayed for several hours or even several days after the marketing information or the reward is received, for example, the user receives a red packet on the nth day, but the user may convert on the N + D days, so that the part of unconverted samples is not a negative sample but a positive sample (called a gray sample) which does not reach the conversion time before the conversion occurs. Due to the fact that user transformation is delayed frequently in a financial scene, and noise information exists in recovered samples, if a traditional gain model is used, due to the fact that the delay of the user transformation is not considered in the basic model G, prediction is inaccurate, prediction of the model for the user transformation is inaccurate, and gain calculation is inaccurate.

In the object identification method provided by the application, the influence of time delay on the conversion possibility of the user is considered, the problem that model prediction is inaccurate due to the fact that the user delays conversion is solved, the accuracy of the prediction of the conversion possibility is improved, and therefore the accuracy of gain calculation is improved.

Experiments prove that the object identification method provided by the application has good effects, the effect of the object identification method is remarkably improved compared with that of the traditional method, disturbance to users who do not convert is reduced, and user experience is improved. The specific effect is shown in table 1, where the number of conversion persons in the table is the total number of conversion persons in the intervention group by the marketing campaign within one week. In effect, compared with the response model modeling, the Uplift modeling method is integrally relatively higher in ROI (return on investment) under the condition that the number of conversion people is unchanged and even is microliter, and in the Uplift modeling method, the effects of TransUplift and TransDelayUplift are better than those of other modeling methods.

TABLE 1 Experimental results

Name of model	Number of transformed persons	Relative lifting	ROI	Relative lifting
					Response model	10389	-	1.151	-
Two Model Approach	10442	0.51％	1.185	2.95％
					One Model Approach	10428	0.38％	1.173	1.91％
TransUplift	10505	1.12％	1.192	3.56％
					TransDelayUplift	10536	1.41％	1.211	5.21％

It should be understood that although the various steps in the flowcharts of fig. 2-9 are shown in order as indicated by the arrows, the steps are not necessarily performed in order as indicated by the arrows. The steps are not limited to being performed in the exact order illustrated and, unless explicitly stated herein, may be performed in other orders. Moreover, at least some of the steps in fig. 2-9 may include multiple steps or multiple stages, which are not necessarily performed at the same time, but may be performed at different times, and the order of performing the steps or stages is not necessarily sequential, but may be performed alternately or alternatively with other steps or at least some of the other steps or stages.

In some embodiments, as shown in fig. 10, there is provided an object recognition apparatus, which may be a part of a computer device using a software module or a hardware module, or a combination of the two, and specifically includes: a training sample set obtaining module 1002, a sample extraction feature obtaining module 1004, a target sample type determining module 1006, a pushed object content conversion degree obtaining module 1008, a reference object content conversion degree obtaining module 1010, a recognition loss value obtaining module 1012, and a trained object recognition model obtaining module 1014, wherein:

a training sample set obtaining module 1002, configured to obtain a training sample set corresponding to an object recognition model to be trained, where the training sample set includes a training sample corresponding to a content push object and a training sample corresponding to a reference object;

a sample extraction feature obtaining module 1004, configured to input training samples in the training sample set into a shared feature extraction network in the object recognition model respectively for feature extraction, so as to obtain sample extraction features corresponding to the training samples;

a target sample type determining module 1006, configured to determine a target sample type corresponding to the sample extraction feature;

a pushed object content conversion degree obtaining module 1008, configured to input the sample extraction feature that the target sample type is the pushed sample type into a pushed object identification network in the object identification model for identification, and identify to obtain a pushed object content conversion degree corresponding to the content pushed object;

a reference object content conversion degree obtaining module 1010, configured to input the sample extraction feature with the target sample type as the reference sample type into a reference object identification network in the object identification model for identification, and identify to obtain a reference object content conversion degree corresponding to the reference object;

an identification loss value obtaining module 1012, configured to obtain a push identification loss value based on a push object content conversion degree corresponding to the content push object, and obtain a reference identification loss value based on a reference object content conversion degree corresponding to the reference object;

and a trained object recognition model obtaining module 1014, configured to perform parameter adjustment on the shared feature extraction network based on the pushed recognition loss value and the reference recognition loss value, and obtain a trained object recognition model based on the adjusted shared feature extraction network.

In some embodiments, the target sample type determination module comprises: the sample label value acquisition unit is used for acquiring the sample label value of the training sample corresponding to the sample extraction characteristic; and the target sample type determining unit is used for determining the target sample type corresponding to the sample extraction feature based on the sample mark value.

In some embodiments, the push object content conversion degree is used to represent a conversion possibility degree of the content push object for the push content, and the module for identifying the loss value comprises: an object identification loss value obtaining unit, configured to perform product calculation on the sample tag value and the content conversion degree of the push object to obtain an object identification loss value corresponding to the content push object; and the pushing identification loss value obtaining unit is used for counting the object identification loss values of the plurality of content pushing objects to obtain the pushing identification loss values.

In some embodiments, the trained object recognition model derivation module comprises: the first parameter adjusting unit is used for carrying out parameter adjustment on the shared feature extraction network based on the push identification loss value to obtain the shared feature extraction network after the push identification loss value is adjusted; the second parameter adjusting unit is used for carrying out parameter adjustment on the shared feature extraction network subjected to the pushing identification loss value adjustment based on the reference identification loss value to obtain a reference identification loss value and the shared feature extraction network subjected to the pushing identification loss value adjustment; and the trained object recognition model acquisition unit is used for extracting a network based on the reference recognition loss value and the shared feature adjusted by the push recognition loss value to obtain a trained object recognition model.

In some embodiments, the first parameter adjusting unit is further configured to determine a push parameter adjustment related value corresponding to the push object identification network based on the push identification loss value; carrying out parameter adjustment on the pushed object identification network based on the pushed parameter adjustment related value to obtain the pushed object identification network after parameter adjustment; determining a first characteristic parameter adjustment related value corresponding to the shared characteristic extraction network based on the pushing identification loss value and the first loss value adjustment direction; and performing parameter adjustment on the shared feature extraction network based on the first feature parameter adjustment related value to obtain the shared feature extraction network with the adjusted push identification loss value.

In some embodiments, the second parameter adjustment unit is further configured to determine a reference parameter adjustment correlation value corresponding to the reference object identification network based on the reference identification loss value; performing parameter adjustment on the reference object identification network based on the reference parameter adjustment related value to obtain the reference object identification network after the parameter adjustment; determining a second characteristic parameter adjustment related value corresponding to the shared characteristic extraction network after the pushing identification loss value is adjusted based on the reference identification loss value and the second loss value adjustment direction; and performing parameter adjustment on the shared feature extraction network after the push identification loss value is adjusted based on the second feature parameter adjustment related value to obtain a reference identification loss value and the shared feature extraction network after the push identification loss value is adjusted.

In some embodiments, the push object identification network comprises a translation likelihood identification network and a likelihood attenuation factor identification network; the push object content conversion degree obtaining module comprises: a total conversion possibility obtaining unit, configured to input the sample extraction feature that the target sample type is the push sample type into a conversion possibility identification network in the push object identification network to perform conversion possibility identification, so as to obtain a total conversion possibility; a likelihood attenuation factor obtaining unit, configured to extract features of a sample in which the target sample type is the push sample type, and input the extracted features into a likelihood attenuation factor recognition network in the push object recognition network to perform attenuation factor recognition, so as to obtain a likelihood attenuation factor; a push object content conversion degree determination unit for determining a push object content conversion degree of the content push object based on the total conversion possibility degree and the possibility degree attenuation factor.

In some embodiments, the push object content conversion degree determination unit is further configured to obtain a possible degree decay time length; determining a subconvertibility degree of the content pushing object based on the attenuation duration of the possibility degree and the attenuation factor of the possibility degree, wherein the subconvertibility degree is used for reflecting the possibility degree of conversion of the object corresponding to the time when the time interval between the content pushing time and the attenuation duration of the possibility degree is the attenuation duration of the possibility degree; and determining the content conversion degree of the push object of the content push object based on the sub-conversion possibility degree and the total conversion possibility degree.

In some embodiments, the training sample set acquisition module comprises: a training push content obtaining unit, configured to obtain training push content; the object dividing unit is used for dividing the training object set to obtain a content pushing object set and a reference object set; the content pushing unit is used for pushing the training pushing content to each content pushing object in the content pushing object set and acquiring the content pushing time of the training pushing content, wherein the reference object in the reference object set shields the automatic pushing training pushing content; the backward operation behavior record acquisition unit is used for acquiring a backward operation behavior record of the content push object aiming at the training push content; a content conversion result determination unit configured to determine a content conversion result of a content push object of the content push objects based on the backward operation behavior record; a content conversion result obtaining unit, configured to obtain, for a reference object in the reference object set, a content conversion result of the reference object after the content push time; and the positive and negative sample obtaining unit is used for taking the training sample corresponding to the object with the converted content conversion result as a positive sample in the training sample set and taking the training sample corresponding to the object with the unconverted content conversion result as a negative sample in the training sample set.

In some embodiments, as shown in fig. 11, there is provided an object recognition apparatus, which may be a part of a computer device using a software module or a hardware module, or a combination of the two, and specifically includes: a candidate object set obtaining module 1102, an object extraction feature obtaining module 1104, a first content conversion possibility obtaining module 1106, a second content conversion possibility obtaining module 1108, a possibility difference obtaining module 1110, and a push object determining module 1112, wherein:

a candidate object set obtaining module 1102, configured to obtain a candidate object set corresponding to a target push content; the candidate object set comprises a plurality of candidate objects;

an object extraction feature obtaining module 1104, configured to input an object sample of the candidate object into a shared feature extraction network of a trained object recognition model to perform feature extraction, so as to obtain an object extraction feature corresponding to the candidate object;

a first content transformation possibility obtaining module 1106, configured to input the object extraction features into a pushed object identification network in the object identification model for identification, and obtain a first content transformation possibility corresponding to the candidate object based on the conversion possibility obtained through identification;

a second content transformation possibility obtaining module 1108, configured to input the object extraction features into a reference object identification network in the object identification model for identification, and obtain a second content transformation possibility corresponding to the candidate object based on the transformation possibility obtained through identification;

a likelihood difference obtaining module 1110, configured to obtain a likelihood difference between a first content conversion likelihood and a second content conversion likelihood corresponding to the candidate object;

a pushed object determining module 1112, configured to filter candidate objects that meet a difference condition from the candidate object set based on a difference in likelihood corresponding to the candidate objects, and use the candidate objects as pushed objects corresponding to the target pushed content; the difference condition includes at least one of a ranking of the energy difference before a preset ranking or a likelihood difference greater than a difference threshold.

In some embodiments, the apparatus further comprises an object recognition model training module comprising: the training sample set acquisition module is used for acquiring a training sample set corresponding to an object recognition model to be trained, wherein the training sample set comprises training samples corresponding to a content push object and training samples corresponding to a reference object; the sample extraction feature obtaining unit is used for respectively inputting the training samples in the training sample set into a shared feature extraction network in the object recognition model for feature extraction to obtain sample extraction features corresponding to the training samples; the target sample type determining unit is used for determining a target sample type corresponding to the sample extraction features; the device comprises a pushed object content conversion degree obtaining unit, a pushed object identification module and a content conversion module, wherein the pushed object content conversion degree obtaining unit is used for inputting the sample extraction characteristics of which the target sample type is the pushed sample type into a pushed object identification network in an object identification model for identification, and identifying to obtain the pushed object content conversion degree corresponding to the content pushed object; the reference object content conversion degree obtaining unit is used for inputting the sample extraction characteristics of which the target sample type is the reference sample type into a reference object identification network in the object identification model for identification, and identifying to obtain the reference object content conversion degree corresponding to the reference object; an identification loss value obtaining unit, configured to obtain a push identification loss value based on a push object content conversion degree corresponding to the content push object, and obtain a reference identification loss value based on a reference object content conversion degree corresponding to the reference object; and the trained object recognition model obtaining unit is used for carrying out parameter adjustment on the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and obtaining a trained object recognition model based on the adjusted shared feature extraction network.

For the specific definition of the object recognition device, reference may be made to the above definition of the object recognition method, which is not described herein again. The modules in the object recognition apparatus may be wholly or partially implemented by software, hardware, and a combination thereof. The modules can be embedded in a hardware form or independent of a processor in the computer device, and can also be stored in a memory in the computer device in a software form, so that the processor can call and execute operations corresponding to the modules.

In some embodiments, a computer device is provided, which may be a terminal, and its internal structure diagram may be as shown in fig. 12. The computer device comprises a processor, a memory, a communication interface, a display screen and an input device which are connected through a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a nonvolatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of an operating system and computer programs in the non-volatile storage medium. The communication interface of the computer device is used for carrying out wired or wireless communication with an external terminal, and the wireless communication can be realized through WIFI, an operator network, NFC (near field communication) or other technologies. The computer program is executed by a processor to implement an object recognition method. The display screen of the computer equipment can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer equipment can be a touch layer covered on the display screen, a key, a track ball or a touch pad arranged on the shell of the computer equipment, an external keyboard, a touch pad or a mouse and the like.

In some embodiments, a computer device is provided, which may be a server, the internal structure of which may be as shown in fig. 13. The computer device includes a processor, a memory, and a network interface connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a nonvolatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operating system and the computer program to run on the non-volatile storage medium. The database of the computer device is used for storing data involved in the object recognition method. The network interface of the computer device is used for communicating with an external terminal through a network connection. The computer program is executed by a processor to implement an object recognition method.

It will be appreciated by those skilled in the art that the configurations shown in fig. 12 and 13 are only block diagrams of portions of configurations relevant to the present application, and do not constitute a limitation on the computer apparatus to which the present application may be applied, and that a particular computer apparatus may include more or fewer components than those shown in the figures, or may combine certain components, or have a different arrangement of components.

In some embodiments, there is also provided a computer device comprising a memory and a processor, the memory having stored therein a computer program, the processor implementing the steps of the above method embodiments when executing the computer program.

In some embodiments, a computer-readable storage medium is provided, in which a computer program is stored which, when being executed by a processor, carries out the steps of the above-mentioned method embodiments.

It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above can be implemented by hardware instructions of a computer program, which can be stored in a non-volatile computer-readable storage medium, and when executed, can include the processes of the embodiments of the methods described above. Any reference to memory, storage, database or other medium used in the embodiments provided herein can include at least one of non-volatile and volatile memory. Non-volatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, or the like. Volatile Memory can include Random Access Memory (RAM) or external cache Memory. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM), for example.

The technical features of the above embodiments can be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the above embodiments are not described, but should be considered as the scope of the present specification as long as there is no contradiction between the combinations of the technical features.

The above-mentioned embodiments only express several embodiments of the present application, and the description thereof is specific and detailed, but not to be understood as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, and these are all within the scope of protection of the present application. Therefore, the protection scope of the present patent shall be subject to the appended claims.

Claims

1. An object recognition method, characterized in that the method comprises:

acquiring a training sample set corresponding to an object recognition model to be trained, wherein the training sample set comprises training samples corresponding to content push objects and training samples corresponding to reference objects;

respectively inputting the training samples in the training sample set into a shared feature extraction network in the object recognition model for feature extraction to obtain sample extraction features corresponding to the training samples;

determining a target sample type corresponding to the sample extraction features;

inputting the sample extraction characteristics of which the target sample type is the push sample type into a push object identification network in the object identification model for identification, and identifying to obtain the content conversion degree of a push object corresponding to the content push object;

inputting sample extraction features of which the target sample type is a reference sample type into a reference object identification network in the object identification model for identification, and identifying to obtain a reference object content conversion degree corresponding to the reference object;

obtaining a pushing identification loss value based on a pushing object content conversion degree corresponding to the content pushing object, and obtaining a reference identification loss value based on a reference object content conversion degree corresponding to the reference object;

and performing parameter adjustment on the shared feature extraction network based on the push identification loss value and the reference identification loss value, and obtaining a trained object identification model based on the adjusted shared feature extraction network.

2. The method of claim 1, wherein the determining the target sample type corresponding to the sample extraction feature comprises:

obtaining a sample mark value of a training sample corresponding to the sample extraction feature;

and determining a target sample type corresponding to the sample extraction feature based on the sample mark value.

3. The method of claim 2, wherein the push object content conversion degree is used to represent a conversion possibility degree of the content push object for push content, and the deriving a push identification loss value based on the push object content conversion degree corresponding to the content push object comprises:

performing product calculation on the sample marking value and the content conversion degree of the push object to obtain an object identification loss value corresponding to the content push object;

and counting the object identification loss values of the plurality of content push objects to obtain the push identification loss values.

4. The method of claim 1, wherein the performing parameter adjustments on the shared feature extraction network based on the pushed recognition loss values and the reference recognition loss values, and wherein obtaining a trained object recognition model based on the adjusted shared feature extraction network comprises:

performing parameter adjustment on the shared feature extraction network based on the push identification loss value to obtain the shared feature extraction network after the push identification loss value is adjusted;

performing parameter adjustment on the shared feature extraction network after the push identification loss value is adjusted based on the reference identification loss value to obtain a reference identification loss value and the shared feature extraction network after the push identification loss value is adjusted;

and obtaining a trained object recognition model based on the reference recognition loss value and the shared feature extraction network adjusted by the push recognition loss value.

5. The method of claim 4, wherein the performing parameter adjustment on the shared feature extraction network based on the pushed recognition loss value to obtain the shared feature extraction network with the pushed recognition loss value adjusted comprises:

determining a push parameter adjustment related value corresponding to the push object identification network based on the push identification loss value;

performing parameter adjustment on the pushed object identification network based on the pushed parameter adjustment related value to obtain a pushed object identification network after parameter adjustment;

determining a first feature parameter adjustment correlation value corresponding to the shared feature extraction network based on the push identification loss value and a first loss value adjustment direction;

and performing parameter adjustment on the shared feature extraction network based on the first feature parameter adjustment related value to obtain the shared feature extraction network with the adjusted push identification loss value.

6. The method of claim 4, wherein performing parameter adjustment on the pushed recognition loss value-adjusted shared feature extraction network based on the reference recognition loss value to obtain a reference recognition loss value and a pushed recognition loss value-adjusted shared feature extraction network comprises:

determining a reference parameter adjustment correlation value corresponding to the reference object identification network based on the reference identification loss value;

performing parameter adjustment on the reference object identification network based on the reference parameter adjustment related value to obtain a reference object identification network after parameter adjustment;

determining a second characteristic parameter adjustment related value corresponding to the shared characteristic extraction network after the push identification loss value is adjusted based on the reference identification loss value and a second loss value adjustment direction;

and performing parameter adjustment on the shared feature extraction network after the push identification loss value is adjusted based on the second feature parameter adjustment related value to obtain a reference identification loss value and the shared feature extraction network after the push identification loss value is adjusted.

7. The method of claim 1, wherein the push object identification network comprises a conversion likelihood identification network and a likelihood attenuation factor identification network; the step of inputting the sample extraction features of which the target sample type is the push sample type into a push object identification network in the object identification model for identification, and the step of identifying and obtaining the content conversion degree of the push object corresponding to the content push object comprises the following steps:

extracting characteristics of a sample with a target sample type as a push sample type, and inputting the characteristics into a conversion possibility identification network in the push object identification network for conversion possibility identification to obtain a total conversion possibility;

extracting characteristics of a sample with a target sample type as a push sample type, and inputting the characteristics into a possible attenuation factor identification network in the push object identification network for attenuation factor identification to obtain a possible attenuation factor;

determining a push object content conversion for the content push object based on the total conversion likelihood and the likelihood attenuation factor.

8. The method of claim 7, wherein determining a push object content conversion degree for the content push object based on the total conversion likelihood and the likelihood attenuation factor comprises:

acquiring possible attenuation duration;

determining a sub-conversion likelihood of the content push object based on the likelihood decay time and the likelihood decay factor; the sub-conversion possibility degree is used for reflecting the possibility degree of conversion of the object corresponding to the time when the time interval between the sub-conversion possibility degree and the content pushing time is the attenuation duration of the possibility degree;

determining a push object content conversion degree of the content push object based on the sub-conversion possibility degree and the total conversion possibility degree.

9. The method according to claim 1, wherein the obtaining of the training sample set corresponding to the object recognition model to be trained comprises:

acquiring training push content;

dividing a training object set to obtain a content pushing object set and a reference object set;

pushing the training push content to each content push object in the content push object set, and acquiring content push time of the training push content, wherein a reference object in the reference object set shields the training push content from being automatically pushed;

acquiring a backward operation behavior record of the content push object aiming at the training push content;

determining a content conversion result of a content push object in the content push objects based on the backward operation behavior record;

for a reference object in the reference object set, obtaining a content conversion result of the reference object after the content push time;

and taking the training sample corresponding to the object with the converted content conversion result as a positive sample in the training sample set, and taking the training sample corresponding to the object with the unconverted content conversion result as a negative sample in the training sample set.

10. An object recognition method, characterized in that the method comprises:

acquiring a candidate object set corresponding to target push content; the candidate object set comprises a plurality of candidate objects;

inputting the object sample of the candidate object into a shared feature extraction network of a trained object recognition model for feature extraction to obtain an object extraction feature corresponding to the candidate object;

inputting the object extraction features into a pushed object recognition network in the object recognition model for recognition, and obtaining a first content conversion possibility corresponding to the candidate object based on the conversion possibility obtained through recognition;

inputting the object extraction features into a reference object recognition network in the object recognition model for recognition, and obtaining a second content conversion possibility corresponding to the candidate object based on the conversion possibility obtained through recognition;

acquiring a possibility difference between a first content conversion possibility and a second content conversion possibility corresponding to the candidate object;

screening candidate objects meeting a difference condition from the candidate object set based on the difference of the corresponding possibility degrees of the candidate objects, and taking the candidate objects as push objects corresponding to the target push content; the difference condition includes at least one of a ranking of the energy difference before a preset ranking or a likelihood difference greater than a difference threshold.

11. The method of claim 10, wherein the step of deriving the trained object recognition model comprises:

inputting the sample extraction characteristics of which the target sample type is the push sample type into a push object identification network in the object identification model for identification, and identifying to obtain a push object content conversion degree corresponding to the content push object;

and performing parameter adjustment on the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and obtaining the trained object recognition model based on the adjusted shared feature extraction network.

12. An object recognition method apparatus, the apparatus comprising:

the training sample set acquisition module is used for acquiring a training sample set corresponding to an object recognition model to be trained, wherein the training sample set comprises training samples corresponding to a content push object and training samples corresponding to a reference object;

a sample extraction feature obtaining module, configured to input training samples in the training sample set into a shared feature extraction network in the object recognition model respectively for feature extraction, so as to obtain sample extraction features corresponding to the training samples;

the target sample type determining module is used for determining a target sample type corresponding to the sample extraction features;

a pushed object content conversion degree obtaining module, configured to input a sample extraction feature that a target sample type is a pushed sample type into a pushed object identification network in the object identification model for identification, and obtain a pushed object content conversion degree corresponding to the content pushed object through identification;

a reference object content conversion degree obtaining module, configured to input the sample extraction features of which the target sample type is the reference sample type into a reference object identification network in the object identification model for identification, and identify to obtain a reference object content conversion degree corresponding to the reference object;

an identification loss value obtaining module, configured to obtain a push identification loss value based on a push object content conversion degree corresponding to the content push object, and obtain a reference identification loss value based on a reference object content conversion degree corresponding to the reference object;

and the trained object recognition model obtaining module is used for carrying out parameter adjustment on the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and obtaining a trained object recognition model based on the adjusted shared feature extraction network.

13. An object recognition apparatus, characterized in that the apparatus comprises:

the candidate object set acquisition module is used for acquiring a candidate object set corresponding to the target push content; the candidate object set comprises a plurality of candidate objects;

an object extraction feature obtaining module, configured to input the object sample of the candidate object into a shared feature extraction network of a trained object recognition model to perform feature extraction, so as to obtain an object extraction feature corresponding to the candidate object;

a first content conversion possibility obtaining module, configured to input the object extraction feature into a pushed object recognition network in the object recognition model for recognition, and obtain a first content conversion possibility corresponding to the candidate object based on the conversion possibility obtained through recognition;

a second content conversion possibility obtaining module, configured to input the object extraction feature into a reference object recognition network in the object recognition model for recognition, and obtain a second content conversion possibility corresponding to the candidate object based on the conversion possibility obtained through recognition;

a likelihood difference obtaining module, configured to obtain a likelihood difference between a first content conversion likelihood and a second content conversion likelihood corresponding to the candidate object;

a pushed object determination module, configured to filter, based on the difference in likelihood corresponding to the candidate object, a candidate object that satisfies a difference condition from the candidate object set, and use the candidate object as a pushed object corresponding to the target pushed content; the difference condition includes at least one of the ranking of the energy difference being before a preset ranking or the likelihood difference being greater than a difference threshold.

14. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor realizes the steps of the method of any one of claims 1 to 11 when executing the computer program.

15. A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the steps of the method according to any one of claims 1 to 11.