Data processing method, model training method, recommendation method and electronic equipment

Abstract

The embodiment of the invention provides a data processing method, a model training method, a recommendation method and electronic equipment, in the data processing method, a target network model comprising a first network model and a second network model is set, the first network model is a network model obtained based on common-domain data training of a target group, and the second network model is a network model obtained based on common-domain data training of the target group; the second network model is a network model obtained by training on the basis of common-domain data and private domain data of part of individuals in the target group, and after attribute information of a user to be matched and attribute information of an object to be matched are input into the target network model, the input data are processed according to the first network model or the second network model, and obtaining a matching result for the to-be-matched user and the to-be-matched object. According to the embodiment, the matching accuracy of the to-be-matched objects of part of individuals can be improved in a targeted manner by effectively utilizing the private domain data.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of artificial intelligence technology, and in particular to a data processing method, a model training method, a recommendation method, and an electronic device. Background Art

[0002] The rapid development of artificial intelligence technology has made various artificial intelligence applications for users more and more extensive. For example, using neural network models to make targeted predictions for users, match objects (including but not limited to products, various physical objects or virtual objects, etc.), and make recommendations, etc.

[0003] However, these neural network models are still unable to meet actual needs. For example, in the prediction scenario, the existing neural network models cannot meet the needs of some end users to further improve the accuracy of prediction results. Summary of the invention

[0004] In view of this, embodiments of the present application provide a data processing method, a model training method, a recommendation method and an electronic device to at least partially solve the above problems.

[0005] According to a first aspect of an embodiment of the present application, a data processing method is provided, comprising: obtaining input data, the input data comprising attribute information of a user to be matched and attribute information of an object to be matched; inputting the input data into a target network model for predicting a matching target object for the user, the target network model comprising a first network model and a second network model, the first network model being a network model trained based on common domain data of a target group, and the second network model being a network model trained based on common domain data and private domain data of some individuals in the target group; processing the input data by means of the first network model or the second network model to obtain a matching result for the user to be matched and the object to be matched.

[0006] According to a second aspect of an embodiment of the present application, a model training method is provided, comprising: obtaining training data for training a target network model, the training data comprising common-domain data samples of a target group and private-domain data samples of some individuals in the target group, the common-domain data samples and the private-domain data samples both comprising attribute information of sample users to be matched and attribute information of sample objects to be matched; based on the common-domain data samples, training a first network model of the target network model, and, based on the common-domain data samples and the private-domain data samples, training a second network model of the target network model.

[0007] According to the third aspect of an embodiment of the present application, a recommendation method is provided, comprising: in response to an input operation at a user terminal of the target individual, obtaining attribute information of an object to be recommended of the target individual; sending the attribute information of the object to be recommended to a target server; the target server is used to process the acquired attribute information of the user to be recommended and the attribute information of the object to be recommended by a data processing method such as the first aspect, to obtain a matching result for the object to be recommended and the user to be recommended; receiving the matching result fed back by the target server, and displaying it.

[0008] According to a fourth aspect of an embodiment of the present application, a data processing device is provided, including: an acquisition module, used to acquire input data, the input data including attribute information of a user to be matched and attribute information of an object to be matched; an input module, used to input the input data into a target network model for predicting a matching target object for the user, the target network model including a first network model and a second network model, the first network model being a network model trained based on common domain data of a target group, and the second network model being a network model trained based on common domain data and private domain data of some individuals in the target group; a prediction module, used to process the input data through the first network model or the second network model to obtain a matching result for the user to be matched and the object to be matched.

[0009] According to a fifth aspect of an embodiment of the present application, a model training device is provided, comprising: a training data acquisition module, used to acquire training data for training a target network model, the training data comprising common domain data samples of a target group and private domain data samples of some individuals in the target group, the common domain data samples and the private domain data samples both comprising attribute information of sample users to be matched and attribute information of sample objects to be matched; a training module, used to train a first network model of the target network model based on the common domain data samples, and to train a second network model of the target network model based on the common domain data samples and the private domain data samples.

[0010] According to the sixth aspect of the embodiment of the present application, a recommendation device is provided, including: an information acquisition module, used to obtain the attribute information of the object to be recommended of the target individual in response to the input operation on the user terminal of the target individual; a sending module, used to send the attribute information of the object to be recommended to the target server; the target server is used to process the acquired attribute information of the user to be recommended and the attribute information of the object to be recommended by the data processing method as described in the first aspect, and obtain a matching result for the object to be recommended and the user to be recommended; a receiving module, used to receive the matching result fed back by the target server and display it.

[0011] According to the seventh aspect of the embodiments of the present application, there is provided an electronic device, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other through the communication bus; the memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform an operation corresponding to the method described in any one of the first to third aspects.

[0012] According to an eighth aspect of the embodiments of the present application, a computer storage medium is provided, on which a computer program is stored. When the program is executed by a processor, the method described in any one of the first to third aspects is implemented.

[0013] According to a ninth aspect of an embodiment of the present application, a computer program product is provided, comprising computer instructions, wherein the computer instructions instruct a computing device to perform operations corresponding to the method described in any one of the first to third aspects.

[0014] The data processing method, model training method, recommendation method and electronic device provided in the embodiments of the present application are configured to include a target network model including a first network model and a second network model, wherein the first network model is a network model trained based on the common domain data of the target group, and the second network model is a network model trained based on the common domain data and the private domain data of some individuals in the target group, so that after the attribute information of the user to be matched and the attribute information of the object to be matched are input into the target network model, the input data can be processed according to the first network model or the second network model to obtain the matching result for the user to be matched and the object to be matched. In this embodiment, a second network model is separately set for some individuals in the target group that provide private domain data, so that more accurate matching prediction can be made for some individuals in the target group that provide private domain data, thereby improving the accuracy of its matching results. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0016] Figure 1 is an exemplary system diagram of a data processing method according to an embodiment of the present application;

[0017] Figure 2 is a flowchart of a data processing method according to an embodiment of the present application;

[0018] Figure 3ASchematic diagram of a preprocessing process of input data for a data processing method according to an embodiment of the present application;

[0019] Figure 3B A schematic diagram of a processing process for prediction results of a data processing method according to an embodiment of the present application;

[0020] Figure 4 A flowchart of the steps of a model training method according to an embodiment of the present application;

[0021] Figure 5A A schematic diagram of the target network model training process of an embodiment of the present application;

[0022] Figure 5B for Figure 5A A schematic diagram of a process for training a second network model in a target network model of the illustrated embodiment;

[0023] Figure 5C for Figure 5A A schematic diagram of a process flow for feature extraction network training in a target network model of the illustrated embodiment;

[0024] Figure 5D A schematic diagram of a processing process of an exemplary target server applicable to an embodiment of the present application;

[0025] Fig. 6A A flowchart of a recommended method according to an embodiment of the present application;

[0026] Figure 6B A schematic diagram of a scenario of a recommended method according to an embodiment of the present application;

[0027] Figure 7 is a structural block diagram of a data processing device according to an embodiment of the present application;

[0028] Figure 8 is a structural block diagram of a model training device according to an embodiment of the present application;

[0029] Fig. 9 is a structural block diagram of a recommendation device according to an embodiment of the present application;

[0030] Fig.10 It is a schematic diagram of the structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0031] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments in the embodiments of the present application should fall within the scope of protection of the embodiments of the present application.

[0032] Before describing the data processing method of the embodiment of the present application, a brief description of the applicable system of the data processing method is first given to facilitate understanding.

[0033] Reference Figure 1 , Figure 1 A schematic diagram of an exemplary system applicable to the data processing method of an embodiment of the present application.

[0034] The system of this embodiment may include a first user terminal, a target server, and a second user terminal, wherein the first user terminal and the second user terminal are both communicatively connected to the target server, and the first user terminal and the second user terminal may be one or more. This embodiment is described by taking the example that the first user terminal and the second user terminal both include multiple terminals.

[0035] The target server can be any appropriate device for storing information, data, programs and / or any other suitable type of content, including but not limited to distributed storage system devices, server cluster devices, computing cloud service cluster devices, etc. In some embodiments, the target server can perform any appropriate function. For example, when the scheme of the embodiment of the present application is implemented by the target server, in some embodiments, the target server can be used to execute the data processing method. As an optional example, in some embodiments, the target server can first obtain input data, and the input data includes attribute information of the user to be matched and attribute information of the object to be matched; the input data is input into a target network model for predicting a matching target object for the user, and the target network model includes a first network model and a second network model, the first network model is a network model trained based on the common domain data of the target group, and the second network model is a network model trained based on the common domain data and the private domain data of some individuals in the target group; the input data is processed by the first network model or the second network model to obtain a matching result for the user to be matched and the object to be matched. In some embodiments, the target server can receive attribute information of the user to be matched sent by the first user terminal, and attribute information of the object to be matched sent by the second user terminal, and after obtaining the matching result for the user to be matched and the object to be matched in the above-mentioned manner, send the matching result to the second user terminal so that the second user terminal displays the matching result.

[0036] In some embodiments, the first user terminal and the second user terminal may include any suitable type of device. For example, in some embodiments, the first user terminal and the second user terminal may include a mobile device, a tablet computer, a laptop computer, a desktop computer and / or any other suitable type of user equipment.

[0037] The following uses an advertising delivery scenario as an example to illustrate the above system. In an advertising delivery scenario, it usually includes users, advertisers, and advertising delivery platforms, where users refer to the recipients of advertisements, advertisers refer to the publishers of advertisements, and advertising delivery platforms are intermediaries connecting users and advertisers, and are used to push advertisements delivered by advertisers to corresponding users.

[0038] In the advertising delivery scenario, in this example, the target server corresponds to the advertising delivery platform, the first user terminal corresponds to the device used by the user, and the second user terminal corresponds to the device used by the advertiser.

[0039] The first user terminal is used to obtain the user's attribute information and send the user's attribute information to the target server; the second user terminal is used to obtain the attribute information of the advertiser's to-be-matched advertisement and send the attribute information of the to-be-matched advertisement to the target server; the target server is deployed with a target network model for predicting matching target advertisements for users. For any received user attribute information, the target network model can be used to predict the user's attribute information and the attribute information of multiple advertisers' to-be-matched advertisements respectively to obtain prediction results for the user and the multiple to-be-matched advertisements. The prediction results may include click-through rate and / or conversion rate. Then, based on the prediction results for the user and the multiple to-be-matched advertisements, a target advertisement matched for the user can be obtained and pushed to the user terminal corresponding to the user; and the matching results of multiple advertisers' to-be-matched advertisements can also be fed back to the corresponding second user terminal.

[0040] It should be noted that the click-through rate refers to the click-through rate of online advertising, that is, the actual number of clicks on the advertisement divided by the number of advertisement impressions. The conversion rate refers to the conversion rate from users clicking on the advertisement to becoming effectively activated or registered or even paying users.

[0041] The specific implementation of the embodiment of the present application is further explained below in conjunction with the accompanying drawings of the embodiment of the present application.

[0042] Reference Figure 2 , Figure 2 1 is a flowchart of a data processing method according to an embodiment of the present application. This embodiment describes the scheme of the embodiment of the present application from the application level. Specifically, the data processing method provided in this embodiment may include the following steps:

[0043] S202: Obtain input data.

[0044] The input data includes attribute information of the user to be matched and attribute information of the object to be matched.

[0045] Specifically, the attribute information of the user to be matched includes information associated with the characteristics of the target object of the user to be matched, for example, the attribute information of the user to be matched may include the user's identity information, interest preference information, behavior information, historical matching records, etc., wherein the user's identity information may include ID, age, gender, etc. The attribute information of the object to be matched includes information associated with the characteristics of the target user of the object to be matched, and the attribute information of the object to be matched may include the object's identification information, content information, category information, usage information, etc., and the identification information may include the object's ID, the ID of the target individual to which the object belongs, etc.

[0046] S204: Input the input data into a target network model for predicting a matching target object for the user.

[0047] Among them, the target network model includes a first network model and a second network model. The first network model is a network model trained based on the common domain data of the target group, and the second network model is a network model trained based on the common domain data and the private domain data of some individuals in the target group.

[0048] Exemplarily, both the first network model and the second network model may be neural network models formed by multiple fully connected layers. Common domain data may refer to shared data obtained from a common database of a target group, and the common domain data includes the following data types: feature data and / or corresponding label data composed of user attribute information and object attribute information; private domain data may refer to privacy data provided by some individuals. The following data types are included in private domain data: feature data and / or corresponding label data composed of user attribute information and object attribute information. Taking the advertising delivery scenario as an example, the target group may refer to all advertisers of the target advertising delivery platform, and some individuals of the target group may refer to target advertisers who provide private domain data on the target advertising delivery platform. Common domain data includes shared data for all advertisers collected by the target advertising delivery platform, and private domain data includes privacy data provided by the target advertiser to the target advertising delivery platform, such as attribute information of advertisements delivered by the target advertiser on other platforms and attribute information of the user, as well as corresponding matching results. The first network model is a network model trained based on the common domain data of the target group. It can be understood that the first network model is a network model trained with the shared data of all advertisers in the advertising delivery platform. Here, the first network model is one; the second network model is a network model trained based on the common domain data and the private domain data of some individuals in the target group. Here, the second network model in the target prediction model can be multiple, and each individual in the partial individuals corresponds to a second network model. For example, the second network model of a certain target advertiser (i.e., a certain individual in the partial individuals) is a network model trained with the shared data corresponding to the target advertiser in the advertising delivery platform and the private domain data provided by the target advertiser. Optionally, the second network model in the target prediction model can be multiple, and each individual in the target group (including some individuals) corresponds to a second network model, wherein the second network model corresponding to each individual in the partial individuals is obtained by training the common domain data corresponding to the individual and the private domain data provided by the individual, and the second network model corresponding to other individuals in the target group except some individuals (i.e., individuals who do not provide private domain data) is obtained by training the common domain data corresponding to the individual.

[0049] S206: Process the input data through the first network model or the second network model to obtain a matching result for the to-be-matched user and the to-be-matched object.

[0050] Specifically, the first network model or the second network model can be selected to perform prediction processing on the input data according to the attribute information of the object to be matched. For example, the attribute information of the object to be matched includes the ID of the target individual to which the object to be matched belongs. According to the ID of the target individual, it can be determined whether the target individual provides private domain data. If the target individual provides private domain data, the second network model corresponding to the target individual can be selected to perform prediction processing on the input data; if the target individual does not provide private domain data, the first network model can be selected to perform prediction processing on the input data. The first network model or the second network model is used to perform prediction processing on the input data to obtain prediction results for the user to be matched and the object to be matched, such as the degree of matching, and the prediction results are used as the matching results of the user to be matched and the object to be matched.

[0051] Furthermore, after respectively obtaining prediction results for multiple users to be matched and multiple objects to be matched, the matching result of each user to be matched and the object to be matched can be determined according to the prediction results of each user to be matched and the multiple objects to be matched, as well as preset screening rules, that is, the matching result can also include a target object to be matched for the user to be matched. For example, the object to be matched with the highest matching degree with the user to be matched can be selected as the target object, and this embodiment does not limit this.

[0052] In this embodiment, a second network model is set for each of the individuals in the target group that provides private domain data, so that more accurate matching predictions can be made for the individuals in the target group that provides private domain data. Compared with the results of matching predictions using the first network model, the prediction results of the second network model can effectively utilize private domain data to specifically improve the matching accuracy of the objects to be matched for these individuals.

[0053] In some optional embodiments, the target group includes a first group and a second group, and the aforementioned part of the individuals belongs to the second group. In this case, the input data is predicted and processed by the first network model or the second network model to obtain a matching result for the user to be matched and the object to be matched, including: if the attribute information of the object to be matched indicates that the object to be matched corresponds to the first group, the input data is processed by the first network model to obtain a matching result for the user to be matched and the object to be matched; if the attribute information of the object to be matched indicates that the object to be matched corresponds to the second group, the input data is processed by the second network model to obtain a matching result for the user to be matched and the object to be matched.

[0054] Exemplarily, the first group includes a plurality of first individuals who have not provided private domain data, and the second group includes a plurality of second individuals who have provided private domain data. In this embodiment, the first group shares a first network model, and each second individual in the second group corresponds to a second network model. If the attribute information of the object to be matched indicates that the object to be matched belongs to the first group, for example, the attribute information of the object to be matched includes the ID of the target individual to which the object to be matched belongs, if the target individual is determined to belong to the first group based on the ID of the target individual, the first network model can be selected to perform prediction processing on the input data; if the target individual is determined to belong to the second group based on the ID of the target individual, the second network model corresponding to the target individual can be selected to perform prediction processing on the input data.

[0055] The matching results of the user to be matched and the object to be matched may include the matching results of each user to be matched and each object to be matched. For example, if the user to be matched is A, and the objects to be matched include three, namely a, b and c, the matching results of the user to be matched A and the objects to be matched may include: the matching results of the user to be matched A and the object to be matched a, the matching results of the user to be matched A and the object to be matched b, and the matching results of the user to be matched A and the object to be matched c, wherein the matching results may include the click-through rate and / or the conversion rate representing the matching degree. Then, the object to be matched with the greatest matching degree in the matching results may be determined as the target object for the user to be matched. For example, in the matching results, the conversion rate of the user to be matched A and the object to be matched a is the first, the conversion rate of the user to be matched A and the object to be matched b is the second, and the conversion rate of the user to be matched A and the object to be matched c is the third, then the object to be matched a is determined to be the target object for the user to be matched A.

[0056] In this embodiment, based on the attribute information of the object to be matched, it is determined whether the object to be matched corresponds to the first group that does not provide private domain data, or the second group that provides private domain data. If the object to be matched corresponds to the first group, the first network model is selected to predict the input data. If the object to be matched corresponds to the second group, the second network model is selected to predict the input data. For the second group that provides private domain data, the second network model is used to predict the matching result of the attribute information of the object to be matched and the attribute information of the user to be matched. The private domain data can be effectively used to improve the matching accuracy of the second group's objects to be matched. In addition, for the first group that does not provide private domain data, and for the second group that provides private domain data, the second network model is used to predict the matching result of the object to be matched and the attribute information of the user to be matched. The second group of private domain data sets the first network model and the second network model respectively, so that the private domain data of the second group is only used to train its corresponding second network model, and the private domain data of some individuals are only used to train their corresponding second network model, instead of using the private domain data of some individuals to train the first network model. Therefore, in the prediction process, the first network model can be used to predict the objects to be matched of the target group, and only the second network model is used to predict the objects to be matched of some individuals. This ensures that the private domain data of some individuals are only used to improve the prediction accuracy of the objects to be matched of some individuals, thereby ensuring the security of the private domain data of some individuals and realizing the privacy isolation of the private domain data of some individuals.

[0057] In some optional implementations, the second network model includes multiple second network models, and the multiple second network models respectively correspond to multiple individuals in the second group. The method of this embodiment also includes: if the attribute information of the object to be matched indicates that the object to be matched corresponds to the second group, obtaining first configuration information for the object to be matched, the first configuration information is used to indicate whether the second network model corresponding to the individual to which the object to be matched belongs is enabled; if the first configuration information indicates that the second network model corresponding to the individual to which the object to be matched belongs is enabled, processing the input data through the second network model to obtain a matching result for the user to be matched and the object to be matched.

[0058] Exemplarily, the first configuration information can be obtained by flexibly configuring according to individual needs during the application process, and the first configuration information is used to indicate whether the second network model corresponding to the individual to which the object to be matched belongs is enabled. For the object to be matched of the target individual who provides private domain data, when the second network model of the target individual is enabled, the second network model is selected for prediction processing; when the second network model of the target individual is not enabled, the first network model is selected for prediction processing. Optionally, the input data can also be predicted and processed by the first network model and the second network model respectively to obtain a first prediction result and a second prediction result. If the first configuration information indicates that the second network model corresponding to the individual to which the object to be matched belongs is enabled, then based on the second prediction result, a matching result for the user to be matched and the object to be matched is obtained; if the first configuration information indicates that the second network model corresponding to the individual to which the object to be matched belongs is not enabled, then based on the first prediction result, a matching result for the user to be matched and the object to be matched is obtained;

[0059] In this implementation, by setting individual first configuration information, it is possible to flexibly choose whether to enable the second network model for prediction based on the first configuration information, thereby achieving flexible pluggable capabilities for the second network model.

[0060] In some optional embodiments, the input data is processed through the first network model to obtain a matching result for the user to be matched and the object to be matched, including: based on the input data, respectively performing probability predictions for indicating that the user to be matched performs a first operation and a second operation on the object to be matched, and obtaining corresponding first operation prediction results and second operation prediction results; based on the first operation prediction results and the second operation prediction results, obtaining a matching result for the user to be matched and the object to be matched.

[0061] Exemplarily, the first network model can predict the probability of the input data indicating that the user to be matched performs a first operation and a second operation on the object to be matched. The first operation of the user to be matched on the object to be matched can be any operation of the user to be matched on the object to be matched, for example, it can be a click operation, and the first operation prediction result can be a click rate. The second operation of the user to be matched on the object to be matched can be any subsequent operation performed by the user to be matched after performing the first operation on the object to be matched, for example, it can be an operation such as registration, purchase, and recharge, and the second operation prediction result can be a conversion rate. Based on the first operation prediction result and the second operation prediction result, a matching result for the user to be matched and the object to be matched is obtained, that is, the matching result can be a click rate and a conversion rate.

[0062] In this embodiment, the first network model can simultaneously predict the first operation prediction result and the second operation prediction result of the user to be matched and the object to be matched, that is, the first network model can simultaneously realize independent prediction of multiple tasks, thereby improving the processing efficiency of prediction for multiple tasks.

[0063] In some optional embodiments, the input data is processed through the second network model to obtain recommendation results for the user and the object to be matched, including: based on the input data, respectively performing probability predictions for indicating that the user to be matched performs a first operation and a second operation on the object to be matched, and obtaining corresponding third operation prediction results and fourth operation prediction results; based on the third operation prediction results and the fourth operation prediction results, obtaining matching results for the user to be matched and the object to be matched.

[0064] Exemplarily, the second network model can respectively predict the probability of the input data indicating that the user to be matched performs a first operation and a second operation on the matching object. The first operation of the user to be matched on the matching object can be any operation of the user to be matched on the matching object, such as a click operation. The third operation prediction result can be a click-through rate. The second operation of the user to be matched on the matching object can be any subsequent operation performed by the user to be matched after the user to be matched performs the first operation on the matching object, such as registration, purchase, recharge and other operations. The fourth operation prediction result can be a conversion rate. Based on the third operation prediction result and the fourth operation prediction result, a matching result for the user to be matched and the matching object is obtained, that is, the matching result can be a click-through rate and a conversion rate.

[0065] In this embodiment, the second network model can also simultaneously predict the third operation prediction result and the fourth operation prediction result of the to-be-matched user and the to-be-matched object, and can also simultaneously realize independent prediction of multiple tasks, thereby improving the processing efficiency of prediction for multiple tasks.

[0066] In some optional embodiments, obtaining input data includes: obtaining initial attribute information of the object to be matched and the target individual to which the object to be matched belongs; replacing the initial attribute information of the object to be matched with the attribute information of the object to be matched according to the preset association relationship between target individuals and / or the association relationship between objects possessed by the target individual; and determining the input data according to the attribute information of the object to be matched and the obtained attribute information of the user to be matched.

[0067] Exemplarily, taking the online commodity sales scenario as an example, the initial attribute information of the object to be matched may be the attribute information of product a placed by commodity seller A, and the target individual to which the object to be matched belongs may be commodity seller A. The preset association relationship between target individuals and / or the association relationship between objects possessed by the target individuals are obtained. For example, the association relationship between target individuals may be the association relationship between commodity seller A and commodity seller B, and the association relationship between objects possessed by the target individuals may be the association relationship between product a of commodity seller A and product b of commodity seller B. The preset association relationship between target individuals and / or the association relationship between objects possessed by the target individuals may be obtained, and then based on the association relationship between target individuals and / or the association relationship between objects possessed by the target individuals, the initial attribute information of the object to be matched is replaced to obtain the attribute information of the object to be matched. For example, part of the attribute information of product a may be replaced with the corresponding information in the product of commodity seller B, or part of the attribute information of product a may be replaced with the corresponding information in product b, and so on. The attribute information of the object to be matched obtained and the attribute information of the user to be matched obtained form the input data.

[0068] For example, refer to Figure 3A , the initial association relationship dictionary can be configured offline in advance. Taking the advertising delivery scenario as an example, the initial association relationship can be determined based on the similarity between advertisers or advertisements. In the offline part, the initial attribute information of the test object used for offline testing is obtained, and the initial attribute information of the test object is replaced according to the initial association relationship dictionary to obtain the attribute information of the test object; the target network model is used for simulation prediction, and the first prediction result is determined based on the initial attribute information of the test object and the attribute information of the test user; the second prediction result is determined based on the attribute information of the test object and the attribute information of the test user, and the first prediction result is compared with the second prediction result to determine whether the second prediction result is better than the first prediction result. If not, the initial association relationship dictionary is adjusted until the second prediction result is better than the first prediction result. After that, it is further determined whether the second prediction result reaches the preset target prediction result. If so, the determined association relationship dictionary can be used online. If not, the above process is continued to be executed in a loop. The association relationship dictionary includes association relationships between target individuals and / or association relationships between objects possessed by target individuals.

[0069] In addition, whether in the offline part or the online part, the initial attribute information of the object to be matched before replacement, the attribute information of the object to be matched after replacement, and the matching results of the prediction processing of the attribute information of the object to be matched after replacement and the attribute information of the user to be matched by the target network model can be counted and the statistical results can be fed back, so that the association relationship dictionary can be adjusted according to the feedback statistical results.

[0070] In this embodiment, the matching results of the objects to be matched with poor matching results can be optimized by replacing the attribute data of the objects to be matched with poor matching results with the attribute data of the objects to be matched with better matching results through the preset association relationships between target individuals and / or the association relationships between objects possessed by the target individuals.

[0071] In some optional embodiments, the input data is processed through the first network model or the second network model to obtain matching results for the user to be matched and the object to be matched, including: processing the input data through the first network model or the second network model to obtain initial prediction results for the user to be matched and the object to be matched; weighting the initial prediction results to obtain matching results for the user to be matched and the object to be matched.

[0072] Exemplarily, the input data is predictively processed through the first network model or the second network model to obtain an initial prediction result for the user to be matched. For example, in the initial prediction result, the conversion rate a1 between the user to be matched and the object to be matched a is first, the conversion rate b1 between the user to be matched and the object to be matched b is second, and the conversion rate c1 between the user to be matched and the object to be matched c is third. According to the preset weight adjustment rule, the conversion rate b1 between the user to be matched and the object to be matched b can be weighted so that the conversion rate b1 is greater than the conversion rate a1. In the final matching result, the conversion rate b1 is first, the conversion rate a1 is second, and the conversion rate c1 is third.

[0073] The preset weight adjustment rules may include weights corresponding to different individuals, and each individual weight may include sub-weights of multiple dimensions. For example, taking the advertising delivery scenario as an example, the weight corresponding to the advertiser may include sub-weights of multiple dimensions such as delivery time period and user source. Figure 3B For a certain individual, multiple pre-set dimensions and corresponding sub-weights are obtained, and then the sub-weights of multiple dimensions are fused and calculated to obtain the weight corresponding to the individual, and the weights corresponding to multiple individuals are combined to generate weight adjustment rules.

[0074] Optionally, statistics may be performed on the initial prediction results of the to-be-matched object before adjustment and the matching results after adjustment, and the statistical results may be fed back, so that the weight adjustment rule may be adjusted according to the fed-back statistical results.

[0075] In this embodiment, after the input data is predicted and processed by the first network model or the second network model to obtain the initial prediction results for the user to be matched and the object to be matched, the initial prediction results can be further weighted and processed to finally obtain the matching results for the user to be matched and the object to be matched. This allows the matching results of the target individual's object to be matched and the user to be matched to be flexibly adjusted according to the personalized needs of the target individual to whom the object to be matched belongs, thereby meeting the personalized needs of the target individual.

[0076] Reference Figure 4 , Figure 4 The following is a flowchart of a model training method according to an embodiment of the present application. This embodiment describes the training process of the target network model in the above data processing method. Specifically, the model training method provided in this embodiment may include the following steps:

[0077] S402, obtaining training data for training the target network model.

[0078] The training data includes common domain data samples of the target group and private domain data samples of some individuals in the target group. Both the common domain data samples and the private domain data samples include attribute information of the sample users to be matched and attribute information of the sample objects to be matched.

[0079] Exemplarily, the attribute information of the sample user to be matched includes information associated with the characteristics of the target sample object of the sample user to be matched, such as identity information, interest preference information, behavior information, historical matching records, etc. of the sample user to be matched, wherein the identity information of the sample user to be matched may include ID, age, gender, etc. The attribute information of the sample object to be matched includes information associated with the characteristics of the target sample user of the sample object to be matched, such as identification information, content information, category information, usage information, etc. of the sample object to be matched, and the identification information may include the ID of the sample object to be matched, the ID of the target individual to which the sample object to be matched belongs, etc.

[0080] The common domain data sample may refer to a shared data sample obtained from a common database of the target group. The common domain data sample includes data types such as sample feature data and / or corresponding label data composed of user attribute information and object attribute information; the private domain data sample may refer to a privacy data sample provided by some individuals. The private domain data sample includes data types such as sample feature data and / or corresponding label data composed of attribute information of the sample user to be matched and attribute information of the sample object to be matched. This embodiment does not limit this. Taking the advertising delivery scenario as an example, the target group may refer to all advertisers of the target advertising delivery platform, and some individuals of the target group may refer to target advertisers who provide private domain data samples on the target advertising delivery platform. Some individuals in the target group may be one or more individuals, each of which is an advertiser who provides private domain data samples. The common domain data sample includes shared data samples for all advertisers collected by the target advertising delivery platform, and the private domain data sample includes privacy data samples provided by the target advertiser to the target advertising delivery platform, such as attribute information of advertisements delivered by the target advertiser on other platforms and attribute information of users, as well as corresponding matching results.

[0081] S404: Based on the common domain data samples, a first network model of the target network model is trained, and based on the common domain data samples and the private domain data samples, a second network model of the target network model is trained.

[0082] Exemplarily, the target network model may include a first network model and a second network model, wherein both the first network model and the second network model may be neural network models formed by multiple fully connected layers. The first network model may be trained by common domain data samples, for example, the first network model may be trained by shared data samples of all advertisers in the advertising delivery platform, where the first network model is one; the second network model may be trained by common domain data samples and private domain data samples, where the second network model in the target prediction model may be multiple, and each individual in the partial individuals corresponds to a second network model, for example, the second network model for the target advertiser may be obtained by training the shared data samples corresponding to the target advertiser in the advertising delivery platform and the private domain data samples provided by the target advertiser.

[0083] Optionally, there may be multiple second network models in the target network model, and each individual in the target group (including some individuals) may correspond to a second network model, wherein the second network model corresponding to each individual in the partial individuals is obtained by training with the common domain data samples corresponding to the individual and the private domain data samples provided by the individual, and the second network models corresponding to other individuals in the target group except some individuals (i.e., individuals who do not provide private domain data) are obtained by training with the common domain data samples corresponding to the individual.

[0084] In some optional implementations, the second network model includes multiple second network models, and the multiple second network models respectively correspond to multiple individuals in the partial individuals. The method of this embodiment also includes: obtaining a private domain data sample of an individual and second configuration information corresponding to the individual; the second configuration information is used to indicate whether the private domain data sample of the individual is applied; if the second configuration information indicates that the private domain data sample of the individual is applied, then based on the private domain data sample of the individual, the second network model corresponding to the individual is trained.

[0085] Here, the second configuration information can be obtained by flexibly configuring according to individual needs during the application process. In this implementation, according to the second configuration information, it can be flexibly selected whether to apply the private domain data sample provided by the individual to train the second network model of the individual.

[0086] In this embodiment, a second network model is set for each of the individuals in the target group that provides the private domain data samples, so that more accurate matching predictions can be made for the individuals in the target group that provide the private domain data samples. Compared with the results of matching predictions using the first network model, the prediction results of the second network model can effectively use the private domain data samples to improve the matching accuracy of the objects to be matched for these individuals in a targeted manner; and, the private domain data of some individuals are only used to train their corresponding second network models, while the private domain data of some individuals are not used to train the first network model. Therefore, in the prediction process, the first network model can be used to predict the objects to be matched for the target group, and only the second network model is used to predict the objects to be matched for some individuals, so that the private domain data of some individuals are only used to improve the prediction accuracy of the objects to be matched for some individuals, thereby ensuring the security of the private domain data of some individuals and realizing the privacy isolation of the private domain data of some individuals.

[0087] In some optional embodiments, the first network model includes a first sub-network corresponding to the first task and a first sub-network corresponding to the second task, and the second network model includes a second sub-network corresponding to the first task and a second sub-network corresponding to the second task;

[0088] Based on the common domain data samples, a first network model of the target network model is trained, and based on the common domain data samples and the private domain data samples, a second network model of the target network model is trained, including:

[0089] Based on the common domain data samples, the first sub-network corresponding to the first task is trained, and the first sub-network corresponding to the second task is trained respectively; the first task is used to indicate the probability prediction of the to-be-matched user performing a first operation on the to-be-matched object; the second task is used to indicate the probability prediction of the to-be-matched user performing a second operation on the to-be-matched object; based on the common domain data samples and the private domain data samples, the second sub-network corresponding to the first task is trained, and the second sub-network corresponding to the second task is trained respectively.

[0090] Exemplarily, the first operation of the to-be-matched user on the matching object can be any operation of the to-be-matched user on the matching object, such as a click operation, and the first task, which is used to indicate the probability prediction of the to-be-matched user performing the first operation on the matching object, can be a click rate; the second operation of the to-be-matched user on the matching object can be any subsequent operation performed by the to-be-matched user after performing the first operation on the matching object, such as registration, purchase, recharge and other operations, and the second task, which is used to indicate the probability prediction of the to-be-matched user performing the second operation on the matching object, can be a conversion rate. Accordingly, the first network model can include a first subnetwork corresponding to the click rate and a first subnetwork corresponding to the conversion rate, and the second network model can include a second subnetwork corresponding to the click rate and a second subnetwork corresponding to the conversion rate. Through the common domain data samples, the first subnetwork corresponding to the first task and the first subnetwork corresponding to the second task can be trained respectively; through the common domain data samples and the private domain data samples, the second subnetwork corresponding to the first task and the second subnetwork corresponding to the second task can be trained respectively.

[0091] In this embodiment, the first network model and the second network model can be trained as multi-task sub-networks respectively, so that the first network model and the second network model can both realize independent prediction of multiple tasks at the same time, which can improve the processing efficiency of prediction for multiple tasks.

[0092] In some optional embodiments, the target network model further includes a first feature extraction network, a second feature extraction network, and a feature fusion network;

[0093] Based on the common domain data samples, a first network model of the target network model is trained, and based on the common domain data samples and the private domain data samples, a second network model of the target network model is trained, including:

[0094] The common domain data samples are subjected to feature extraction through a first feature extraction network to obtain first sample feature data; the private domain data samples are subjected to feature extraction through a second feature extraction network to obtain second sample feature data; the first sample feature data and the second sample feature data are fused through a feature fusion network to obtain fused sample feature data; based on the first sample feature data, the first network model is subjected to supervised training; based on the fused sample feature data, the second network model is subjected to supervised training.

[0095] Exemplarily, the target network model may also include a first feature extraction network, a second feature extraction network, and a feature fusion network, wherein the first feature extraction network, the second feature extraction network, and the feature fusion network may all be composed of multiple fully connected layers. The first feature extraction network may be used to extract features from the common domain data samples to obtain first sample feature data, which may be multidimensional features; and the second feature extraction network may be used to extract features from the private domain data samples to obtain second sample feature data, which may be one-dimensional features; finally, the first sample feature data and the second sample feature data may be fused through the feature fusion network to obtain fused sample feature data, for example, the first sample feature data and the second sample feature data may be multiplied, etc. This embodiment does not limit this.

[0096] Accordingly, supervised training is performed on the first network model through the first sample feature data, and the label data in the common domain data sample can be used for training the first network model. Supervised training is performed on the second network model through the fusion sample feature data, and the label data in the private domain data sample can be used for training the first network model.

[0097] In this embodiment, in the target network model, a corresponding first feature extraction network is provided for the common domain data samples, and a corresponding second feature extraction network is provided for the private domain data samples, so as to realize separate processing of feature extraction of common domain data samples and private domain data samples, so as to facilitate the use of common domain data samples and private domain data samples to train different network models in subsequent training process.

[0098] In an optional implementation, the second network model is trained based on the fused sample feature data, including: performing supervised training on the second network model according to the fused sample feature data in combination with the first prediction result of the first network model.

[0099] Specifically, when training the second network model, the second network model can be trained in combination with the fused sample feature data, the label data in the private domain data sample, and the first prediction result of the first network model, so that the parameters of the second network model can be corrected based on the first prediction result of the first network model to improve the accuracy of the model parameters of the second network model.

[0100] In an optional implementation, the training method of this embodiment further includes: training the first feature extraction network based on the common domain data samples; and training the second feature extraction network based on the private domain data samples.

[0101] Specifically, the first feature extraction network is used to extract features from common domain data samples, and the second feature extraction network is used to extract features from private domain data samples. Therefore, in the training phase, the first feature extraction network is trained using common domain data samples respectively; and the second feature extraction network is trained based on private domain data samples, which can specifically improve the accuracy of feature extraction of the first feature extraction network and the second feature extraction network.

[0102] In some optional embodiments, the first feature extraction network and the second feature extraction network may also be divided into different sub-networks according to different tasks. For example, the first feature extraction network may include a first feature extraction sub-network corresponding to a first task (such as click-through rate) and a first feature extraction sub-network corresponding to a second task (such as conversion rate), and the second feature extraction network may include a second feature extraction sub-network corresponding to the first task (such as click-through rate) and a second feature extraction sub-network corresponding to the second task (such as conversion rate).

[0103] When performing feature extraction on the common domain data samples through the first feature extraction network, feature extraction can be performed on the common domain data samples through the first feature extraction subnetwork corresponding to the first task to obtain first subsample feature data corresponding to the first task; and feature extraction can be performed on the common domain data samples through the first feature extraction subnetwork corresponding to the second task to obtain first subsample feature data corresponding to the second task, that is, the first sample feature data can be obtained.

[0104] Similarly, when performing feature extraction on private domain data samples through the second feature extraction network, feature extraction can be performed on the private domain data samples through the second feature extraction sub-network corresponding to the first task to obtain second sub-sample feature data corresponding to the first task; and, feature extraction can be performed on the private domain data samples through the second feature extraction sub-network corresponding to the second task to obtain second sub-sample feature data corresponding to the second task, that is, second sample feature data can be obtained.

[0105] Finally, when the first sample feature data and the second sample feature data are fused through the feature fusion network, the first sub-sample feature data and the second sub-sample feature data corresponding to the first task can be fused through the feature fusion network to obtain the first fused sample feature data; and the first sub-sample feature data and the second sub-sample feature data corresponding to the second task can be fused through the feature fusion network to obtain the second fused sample feature data; that is, the first fused sample feature data and the second fused sample feature data constitute the fused sample feature data.

[0106] In this embodiment, the first network model includes a first subnetwork corresponding to the first task and a first subnetwork corresponding to the second task, and the second network model includes a second subnetwork corresponding to the first task and a second subnetwork corresponding to the second task. Therefore, the first fused sample feature data can be respectively input into the first subnetwork and the second subnetwork corresponding to the first task, and the second fused sample feature data can be respectively input into the first subnetwork and the second subnetwork corresponding to the second task, thereby realizing independent training of feature extraction networks and network models for multiple tasks, and the training processes of feature extraction networks and network models for different tasks can be carried out in parallel.

[0107] In addition, the target network model of this embodiment is an end-to-end network model, which includes a first network model to be trained, a second network model, a first feature extraction network, and a second feature extraction network. The above multiple sub-models are trained in parallel during the training process and do not need to be trained separately and independently, which can save training resources and improve training efficiency.

[0108] Reference Figure 5A , Figure 5A The following is a schematic diagram of the target network model training process of the embodiment of the present application. Figure 5A The target network model training process provided in the embodiment of the present application is explained as follows:

[0109] See also Figure 5A , the target network model includes a network model, a feature extraction network and a feature fusion network; exemplarily, the network model includes a first network model and a second network model, the first network model includes CM0, wherein CM0 refers to CMU0 stored in the form of a vector, CMU0 is CM0 stored in the form of a neural network model, and conversion can be performed between CM0 and CMU0. CMU0 can be divided into CMU according to different tasks 01 、CMU 02 and CMU 03 , among which, CMU 01 Corresponding to the first task, such as click-through rate, CMU 02 and CMU 03Corresponding to the second task, such as conversion rate, the conversion rate can be divided into conversion rate A and conversion rate B according to the conversion type (such as conversion magnitude), that is, CMU 02 Corresponding conversion rate A, CMU 03 The corresponding conversion rate is B. Similarly, the second network model includes CM i , among which, CM i and CMU i Conversion can be performed between them. i According to different tasks, it can be divided into CMU i1 、CMU i2 and CMU i3 , among which, CMU i1 Corresponding to the first task, such as click-through rate, CMU i2 and CMU i3 Corresponding to the second task, such as conversion rate, the conversion rate can be divided into conversion rate A and conversion rate B according to the conversion type (such as conversion magnitude), that is, CMU i2 Corresponding conversion rate A, CMU i3 Corresponding conversion rate B.

[0110] Here, CM0 corresponds to the target group, CM i It can correspond to a target individual in the target group who provides private domain data samples, such as advertiser A who provides private domain data samples, or it can correspond to any individual in the target group. According to the information of the individual contained in the input training data, such as the individual ID, the CM corresponding to the individual ID can be selected from the preset second network model set. i , i.e. Selected FlattenedCM i , the second network model set includes n CM i , such as CM1, CM2…CM n , the obtained CM i Perform the conversion, namely UnfoldCM i , get the corresponding CMU i1 、CMU i2 and CMU i3 In addition, when the training data is obtained, CM0 is obtained, and the obtained CM0 is converted to obtain the corresponding CMU 01 、CMU 02 and CMU 03 .

[0111] The feature extraction network includes a first feature extraction network and a second feature extraction network. The first feature extraction network can be divided into a first feature extraction sub-network CTR corresponding to the first task and a first feature extraction sub-network CVR corresponding to the second task according to different tasks. The first feature extraction sub-network CVR corresponding to the second task can be divided into CVR according to the conversion type. A and CVR B Similarly, the second feature extraction network can be divided into the second feature extraction subnetwork FDN corresponding to the first task according to different tasks. ctr The second feature extraction subnetwork FDN corresponding to the second task cvr , the second feature extraction subnetwork FDN corresponding to the second task cvr Can be divided into FDN according to conversion type cvrA and FDN cvrB .

[0112] The feature fusion network includes scale 0 / 1 .

[0113] When training the target network model, training data is obtained, wherein the training data may include common domain data samples of the target group and private domain data samples of some individuals in the target group. The common domain data samples may refer to shared data samples obtained from a common database of the target group. The common domain data samples include data types such as: sample feature data and / or corresponding label data composed of user attribute information and object attribute information; private domain data samples may refer to privacy data samples provided by an individual. The private domain data samples include data types such as: sample feature data and / or corresponding label data composed of user attribute information and object attribute information, which is not limited in this embodiment.

[0114] For common domain data samples, the following operations can be performed: After embedding the common domain data samples, FeaturesEmbedding is input into the first feature extraction network CTR, CVR A and CVR B , output multi-dimensional features through the first feature extraction network; input the multi-dimensional features into the feature fusion network scale 0 / 1 ;

[0115] For private domain data samples, the following operations can be performed: After embedding the private domain data samples, SpecificEmbedding is input into the second feature extraction network FDN ctr 、FDN cvrA and FDN cvrB , output one-dimensional features through the second feature extraction network; input the one-dimensional features into the feature fusion network scale 0 / 1 ;

[0116] Through feature fusion network scale 0 / 1 The multi-dimensional features output by the first feature extraction network are fused with the one-dimensional features output by the second feature extraction network to output high-dimensional features h ctr 、h cvrA and h cvrB .

[0117] All h ctr 、h cvrA and h cvrB Enter the corresponding CMU in the first network model respectively 01 、CMU 02 and CMU 03 Output the corresponding first prediction results pCTR, pCVR A and pCVR B , based on the first prediction results pCTR, pCVR A 、pCVR B And the corresponding label data in the co-domain data sample, calculate the loss value loss main And loss value loss main-FDN , according to the loss value loss main Adjust the model parameters of the first feature extraction network and the first network model according to the loss value loss main-FDN Adjust the model parameters of the second feature extraction network.

[0118] According to the information of individuals contained in the training data, h corresponding to the individual ctr 、h cvrA and h cvrB Input the CMU of the second network model corresponding to each individual i1 、CMU i2 and CMU i3 Output the corresponding second prediction results pCTR, pCVR A and pCVR B , based on the second prediction results pCTR, pCVR A 、pCVR B And the corresponding label data in the private domain data sample, calculate the loss value loss CM-FDN And loss value loss CM , according to the loss value loss CM Adjust the model parameters of the second network model corresponding to the individual providing the private domain data sample, based on the second prediction results pCTR, pCVR A 、pCVR B And the corresponding label data in the co-domain data sample, calculate the loss value loss CM-FDN And loss value loss CM , according to the loss value lossCM Adjust the model parameters of the second network model corresponding to individuals who do not provide private domain data samples, according to the loss value loss CM-FDN Adjust the model parameters of the second feature extraction network.

[0119] In the embodiment of the present application, there is no limitation on the loss function used to calculate the loss value, and it can be customized according to the actual situation. For example, a logarithmic loss function, a cross entropy loss function, a mean square error loss function, etc. can be used.

[0120] See also Figure 5B , Figure 5B for Figure 5A A schematic diagram of the process of training a second network model in a target network model of the illustrated embodiment.

[0121] In this embodiment, the training process of the second network model is described by taking the case where the input training data includes sample feature data of the common domain data sample and the label data used for supervised training is the label data in the private domain data sample as an example:

[0122] After embedding the sample feature data of the common domain data samples, the features are input into the first feature extraction network CTR and CVR. A and CVR B , output multi-dimensional features through the first feature extraction network; here the first feature extraction network CTR, CVR A and CVR B , which can be tower networks. The multi-dimensional features output by the first feature extraction network are respectively input into the corresponding CMUs in the first network model (not shown in the figure) 01 、CMU 02 and CMU 03 , output the corresponding first prediction results pCVR, pCVR A and pCVR B .

[0123] The multi-dimensional features output by the first feature extraction network and the first prediction results pCTR, pCVR A and pCVR B Enter the CMU in the second network model respectively i1 、CMU i2 and CMU i3 Output the corresponding second prediction results pCTR, pCVR A and pCVR B Based on the second prediction results pCVR, pCVR A 、pCVR B, the labeled data in the private domain data sample and the labeled data in the common domain data sample, and calculate the loss value loss CM , according to the loss value loss CM Adjust model parameters of the first feature extraction network and the second prediction network.

[0124] It should be noted that this embodiment determines which sub-network model parameters in the target network model to adjust based on the set mask. For example, the CVR can be adjusted based on the conversion type convert. A or CVR B Model parameters; CMU can be adjusted based on label data label i or the model parameters of CMU0; in addition, the weight weight of the model parameter can also be adjusted, and the adjustment dimension of the weight can include delay compensation delay and multiple tasks multi, which is not limited in this embodiment. The conversion type, label data and other information can be obtained from the input data through automatic routing, so that the corresponding mask can be generated according to the obtained information, so as to achieve targeted model parameter adjustment when adjusting the model parameters of each sub-network in the target network model.

[0125] See also Figure 5C , Figure 5C for Figure 5A A schematic diagram of the process of training a feature extraction network in a target network model of the illustrated embodiment.

[0126] In this embodiment, the training data input includes sample feature data of common domain data samples and sample feature data of private domain data samples, and the label data used for supervised training is the label data in the common domain data samples as an example to illustrate the training process of the feature extraction network:

[0127] For the common domain data samples, the following operations can be performed: After embedding the sample feature data of the common domain data samples, the features are input into the first feature extraction network CTR, CVR A and CVR B , where the first feature extraction network CTR, CVR A and CVR B , can be tower networks. Output multidimensional features through the first feature extraction network; input the multidimensional features into the feature fusion network convat; input the multidimensional features output by the first feature extraction network into the corresponding CMU in the first network model (not shown in the figure) 01 、CMU 02 and CMU 03 Output the corresponding first prediction results pCTR, pCVR A and pCTRB .

[0128] For private domain data samples, the following operations can be performed: After embedding the sample feature data of the private domain data sample, the sample feature data is input into the second feature extraction network FDN. cvrA and FDN cvrB , output one-dimensional features through the second feature extraction network; input the one-dimensional features into the feature fusion network concat. The example of this embodiment does not set the second feature extraction network FDN for click rate ctr .

[0129] The multi-dimensional features output by the first feature extraction network are fused with the one-dimensional features output by the second feature extraction network through the feature fusion network concat to output high-dimensional features h and h respectively. cvrA and h cvrB (not shown in the figure).

[0130] h cvrA and h cvrB Input the corresponding CMU in the first network model (not shown in the figure) respectively 02 and CMU 03 Output the corresponding third prediction result pCVR A and pPVR B ; Based on the first prediction result, the third prediction result and the label data in the common domain data sample, the loss value loss is calculated, and the model parameters of the first feature extraction network, the second feature extraction network and the first network model are adjusted according to the loss value loss.

[0131] Reference Figure 5D , Figure 5D The present invention is a schematic diagram of a processing procedure of an exemplary target server applicable to an embodiment of the present application.

[0132] As shown in the figure, the target server may include an online service module, a storage / model module and a data interface module.

[0133] Taking the advertising delivery scenario as an example, the data interface module can obtain training data for model training from the advertiser terminal through the external data interface. The training data may include sample feature data and label data. The sample feature data includes user attribute information and advertisement attribute information. The label data includes click-through rate and / or conversion rate. Afterwards, the training data is preprocessed to obtain preprocessed data samples that meet the standardization requirements of model training; the preprocessed data samples may include common domain data samples, private domain data samples and assist labels, wherein the common domain data samples include sample feature data and corresponding label data collected by the target advertising delivery platform for all advertisers, for example, feature data and corresponding label data composed of users and advertisements in the target advertising delivery platform; the private domain data samples include privacy sample feature data and corresponding label data provided by the target advertiser in the target advertising delivery platform, for example, attribute information of advertisements and user attribute information delivered by the target advertiser on other platforms, and corresponding matching results; the assist label may be label data corresponding to the sample feature data in the common domain data sample provided by the target advertiser, for example, feature data composed of users and advertisements on the target advertising delivery platform, and label data corresponding to the feature data provided by the target advertiser on other platforms. The data interface module may also obtain attribute information of the advertisement to be matched from the advertiser terminal through the external data interface, and send the attribute information of the advertisement to be matched to the storage / model module for storage.

[0134] The target network model is deployed in the target server, and the target network model is trained by the common domain data samples and private domain data samples obtained by the data interface module. The assist model can also be deployed in the target server, and the assist model is trained by the assist label obtained by the data interface module. The assist model is a prediction model for predicting conversion rate and / or click-through rate. The model structure and parameters of the assist model are not limited in this embodiment.

[0135] The online service module can obtain attribute information of the user to be matched from the user terminal, and send the attribute information of the user to be matched to the query control unit, which is used to call the data processing model from the storage / model module, such as the assist model and / or the target network model, and obtain the pre-stored attribute information of the advertisement to be matched.

[0136] The received attribute information of the user to be matched and the attribute information of the advertisement to be matched can be predicted by the target network model to obtain the first matching result for the user to be matched and the object to be matched, that is, result a; the received attribute information of the user to be matched and the attribute information of the advertisement to be matched can also be predicted by the assist model to obtain the second matching result of the user to be matched and the object to be matched, that is, result b. The matched advertisement is pushed to the corresponding user terminal according to result a or result b. Here, a better result can be selected from the selection of result a and result b as the final result for advertisement push. For example, result a includes conversion rate a1, and result b includes conversion rate b1. If conversion rate a1 is greater than conversion rate b1, result a is selected as the final result. If conversion rate a1 is less than conversion rate b1, result b is selected as the final result. Optionally, result a and result b can also be weighted to obtain result c, and result c is used as the final result for advertisement push, wherein the weights corresponding to result a and result b can be flexibly set by those skilled in the art according to actual conditions, and this embodiment does not limit this.

[0137] The obtained result a and result b can be counted, for example, by a preset time period, and the obtained statistical results can be fed back to the advertiser terminal through the external data interface. Furthermore, for the attribute information of the same to-be-matched user and the attribute information of the to-be-matched advertisement, result a and result b can be obtained respectively, and the result a and result b can be compared, and the comparison result can be fed back to the advertiser terminal.

[0138] See also Fig. 6A , Fig. 6A 1 is a flowchart of a recommendation method according to an embodiment of the present application. Specifically, the recommendation method provided in this embodiment can be applied to a user terminal of a target individual, such as the second user terminal in the aforementioned exemplary system, and the recommendation method may include the following steps:

[0139] S602: In response to an input operation at a user terminal of the target individual, obtain attribute information of an object to be recommended of the target individual.

[0140] Exemplarily, the user terminal of the target individual (e.g., the second user terminal) may provide an information input interface for the target individual to input or upload attribute information of the object to be recommended, such as attribute information of the advertisement to be recommended. When the second user terminal detects the input operation, the attribute information of the object to be recommended provided by the target individual may be obtained.

[0141] S604, sending the attribute information of the object to be recommended to the target server;

[0142] The target server is used to process the acquired attribute information of the user to be recommended and the attribute information of the object to be recommended by the data processing method as described in the first aspect, and obtain a matching result between the attribute information of the object to be recommended and the user to be recommended.

[0143] The recommendation method of this embodiment is applied to the user terminal of the target individual, such as the second user terminal of the aforementioned exemplary system terminal, and the attribute information of one or more objects to be recommended can be sent to the target server through the second user terminal. The target server is deployed with a target network model, and the target server stores the received attribute information of the objects to be recommended. When the target server receives the attribute information of one or more users to be recommended, the attribute information of the users to be recommended and the attribute information of the objects to be recommended are predicted and processed by the aforementioned data processing method to obtain the prediction results of the attribute information of the objects to be recommended and the users to be recommended, such as the matching degree, and the prediction results can be fed back to the second user terminal as the matching results. Further, after respectively obtaining the prediction results for multiple users to be recommended and each object to be recommended, the matching results of each object to be recommended and the user to be recommended can be determined according to the prediction results of the users to be recommended and the objects to be recommended, and the preset screening rules. The matching results can also include the target users matched with the objects to be recommended. For example, the preset screening rules can be to select the users to be recommended whose matching degree with the objects to be recommended exceeds the preset threshold as the target users, and this embodiment does not limit this.

[0144] In an optional implementation, the target server may collect statistics on the matching results of the recommended objects at a preset frequency, obtain statistical results, and feed back the statistical results.

[0145] S604: Receive the matching result fed back by the target server and display it.

[0146] Exemplarily, the display method may be in the form of a chart or text, which is not limited in this embodiment.

[0147] In an optional implementation, the target server may also receive and display the statistical results of the matching results. Specifically, the target server may count the matching results of the recommended object at a preset frequency, and feed back the statistical results to the second user terminal corresponding to the recommended object for display. The preset frequency may be flexibly set by those skilled in the art, such as every hour, every day, every week, etc., and this embodiment does not limit this.

[0148] In this embodiment, the target individual can input the attribute information of the object to be recommended in the user terminal, and then send the attribute information of the object to be recommended to the target server through the user terminal. The target server is deployed with a target network model, and the target network model includes a second network model set for some individuals who provide private domain data and used separately. When the target individual belongs to some individuals who provide private domain data, the second network model can be used to make a more accurate matching prediction for the target individual's object to be recommended. Compared with the result of matching prediction using the first network model, the prediction result of the second network model can effectively use private domain data to improve the matching accuracy of the target individual's object to be recommended.

[0149] In some optional embodiments, after receiving the matching results fed back by the target server, the method of this embodiment also includes: in response to a first query operation on the user terminal of the target individual, obtaining click data and / or conversion data of the object to be recommended, as well as the matching results of the object to be recommended fed back by the target server; analyzing the matching results of the object to be recommended and the first operation data and / or the second operation data, obtaining the analysis results, and displaying them.

[0150] Exemplarily, the target individual can query the recommendation conversion result for the object to be recommended. For example, the second user terminal can provide a query interface. When the first query operation of the query interface is monitored, such as clicking the first query button, the first operation data and / or the second operation data of the object to be recommended in the preset time period can be obtained, as well as the matching result of the object to be recommended fed back by the target server. Here, the preset time period can be one day, one week, etc., which is not limited in this embodiment. The first operation data can be click data, and the click data can include the user to be recommended who performs a click operation on the object to be recommended. The second operation data can be conversion data, and the conversion data can include the user to be recommended who performs a conversion operation on the object to be recommended. The conversion operation can be an operation such as purchase, registration, and recharge. Afterwards, the matching result of the object to be recommended and the click data and / or the conversion data are analyzed to obtain the analysis result. The analysis result can include the number of recommendations of the object to be recommended and the corresponding first set of users to be recommended, the number of clicks that complete the click operation in the first set of users to be recommended and the corresponding first set of click users, and the number of conversions that complete the conversion operation in the first set of click users and the corresponding first set of conversion users.

[0151] In this embodiment, a query function for the recommendation conversion results of the recommended object can be provided to the target individual. By analyzing the matching results, click data and / or conversion data of the recommended object, the analysis results are obtained and displayed, making it convenient for the target individual to monitor the recommendation status of the recommended object.

[0152] In some optional embodiments, the method of this embodiment also includes: in response to a second query operation on the user terminal of the target individual, obtaining a reference matching result for the object to be recommended fed back by the target server, the reference matching result being obtained by the target server through processing the attribute information of the user to be recommended and the attribute information of the object to be recommended by a preset network model; analyzing the reference matching result and the first operation data and / or the second operation data to obtain a reference analysis result; and displaying the analysis result and the reference analysis result.

[0153] Exemplarily, the preset network model may be the assist model in the aforementioned embodiment, and the reference matching result may be the matching result obtained by processing the attribute information of the user to be recommended and the attribute information of the object to be recommended by the assist model. Similarly, the reference matching result and the first operation data and / or the second operation data may be analyzed to obtain a reference analysis result, which may include the number of recommendations of the object to be recommended and the corresponding second set of users to be recommended, the number of clicks completed in the second set of users to be recommended and the corresponding second set of click users, and the number of conversions completed in the second set of click users and the corresponding second set of conversion users. Afterwards, the analysis result and the reference analysis result may be displayed in the same interface.

[0154] In this embodiment, the target individual can simultaneously query the recommendation status corresponding to the matching of the recommended object through the target network model, and the recommendation status corresponding to the matching of the recommended object through the preset network model. The target individual can understand the recommendation effects of different recommendation methods, and then can select the applicable recommendation method according to the recommendation effect corresponding to the target network model and the recommendation effect corresponding to the preset network model.

[0155] Reference Figure 6B , Figure 6B The present invention is a scenario diagram of a recommended method according to an embodiment of the present application.

[0156] The recommended method of this embodiment is applied to Figure 1 The second user terminal in the exemplary system, taking the online commodity sales scenario as an example, the second user terminal corresponds to the commodity seller M, the target server may be a server of the online sales platform, and the target server may execute the aforementioned data processing method.

[0157] The commodity seller can send the attribute information of the object to be recommended (for example, product A, product B, etc.) to the target server through the second user terminal. The target server will store the received attribute information of the object to be recommended. The target server can also receive the attribute information of the user to be recommended (for example, commodity buyer N). When the target server receives the attribute information of commodity buyer N, the attribute information of commodity buyer N and the attribute information of the object to be recommended are predicted and processed by the data processing method as described above, and the prediction results of product A and commodity buyer N, and the prediction results of product B and commodity buyer N are obtained respectively. For example, the prediction results may include the prediction results of commodity buyer N for Based on the likes of product A and product B, the prediction result can be fed back to the display interface of the second user terminal for display; further, the target product recommended to the commodity purchaser N can be determined according to the likes of the commodity purchaser N for product A and product B respectively. For example, if the target product is product A, the recommendation of product A to the commodity purchaser N can also be fed back to the display interface of the second user terminal as a matching result for display; further, the number of times product A and product B are recommended can be counted according to a preset frequency, for example, every hour, and the statistical results can also be fed back to the display interface of the second user terminal for display. For example, the statistical results can be displayed in the form of a line graph. It should be noted that any one of the above matching results, or any combination of multiple ones, can also be selected for feedback and display, and this embodiment does not limit this.

[0158] See also Figure 7 , Figure 7 1 is a structural block diagram of a data processing device according to an embodiment of the present application. The data processing device provided by the embodiment of the present application includes:

[0159] The acquisition module 702 is used to acquire input data, where the input data includes attribute information of the user to be matched and attribute information of the object to be matched.

[0160] An input module 704 is used to input input data into a target network model for predicting a matching target object for a user. The target network model includes a first network model and a second network model. The first network model is a network model trained based on the common domain data of the target group, and the second network model is a network model trained based on the common domain data and private domain data of some individuals in the target group.

[0161] The prediction module 706 is used to process the input data through the first network model or the second network model to obtain a matching result for the to-be-matched user and the to-be-matched object.

[0162] In some optional embodiments, the target group includes a first group and a second group, and some individuals belong to the second group;

[0163] The prediction module 706 is also used for: if the attribute information of the object to be matched indicates that the object to be matched corresponds to the first group, then the input data is processed through the first network model to obtain a matching result for the user to be matched and the object to be matched; if the attribute information of the object to be matched indicates that the object to be matched corresponds to the second group, then the input data is processed through the second network model to obtain a matching result for the user to be matched and the object to be matched.

[0164] In some optional embodiments, the prediction module 706 is also used to: based on the input data, respectively perform probability predictions for indicating that the user to be matched performs a first operation and a second operation on the object to be matched, and obtain corresponding first operation prediction results and second operation prediction results; based on the first operation prediction results and the second operation prediction results, obtain a matching result for the user to be matched and the object to be matched.

[0165] In some optional embodiments, the prediction module 706 is also used to: based on the input data, respectively perform probability predictions for indicating that the user to be matched performs a first operation and a second operation on the object to be matched, and obtain corresponding third operation prediction results and fourth operation prediction results; based on the third operation prediction results and the fourth operation prediction results, obtain a matching result for the user to be matched and the object to be matched.

[0166] In some optional embodiments, the acquisition module 702 is also used to: obtain initial attribute information of the object to be matched and the target individual to which the object to be matched belongs; replace the initial attribute information of the object to be matched with the attribute information of the object to be matched according to the preset association relationship between target individuals and / or the association relationship between objects possessed by the target individual; determine the input data according to the attribute information of the object to be matched and the obtained attribute information of the user to be matched.

[0167] In some optional embodiments, the prediction module 706 is also used to: process the input data through the first network model or the second network model to obtain initial prediction results for the user to be matched and the object to be matched; and perform weight adjustment on the initial prediction results to obtain matching results for the user to be matched and the object to be matched.

[0168] The data processing device of the embodiment of the present application is used to implement the corresponding data processing method in the aforementioned method embodiment, and has the beneficial effects of the corresponding method embodiment, which will not be repeated here. In addition, the functional implementation of each module in the data processing device of the embodiment of the present application can refer to the description of the corresponding part in the aforementioned method embodiment, which will not be repeated here.

[0169] Reference Figure 8 , Figure 8 The following is a structural block diagram of a model training device according to an embodiment of the present application. The model training device provided in the embodiment of the present application includes:

[0170] The training data acquisition module 802 is used to acquire training data for training the target network model. The training data includes common domain data samples of the target group and private domain data samples of some individuals in the target group. Both the common domain data samples and the private domain data samples include attribute information of the sample users to be matched and attribute information of the sample objects to be matched.

[0171] The training module 804 is used to train a first network model of the target network model based on the common domain data samples, and to train a second network model of the target network model based on the common domain data samples and the private domain data samples.

[0172] In an optional implementation, the first network model includes a first sub-network corresponding to the first task and a first sub-network corresponding to the second task, and the second network model includes a second sub-network corresponding to the first task and a second sub-network corresponding to the second task;

[0173] The training module 804 is also used to: based on the common domain data samples, train the first sub-network corresponding to the first task and the first sub-network corresponding to the second task respectively; the first task is used to indicate the probability prediction of the to-be-matched user performing the first operation on the to-be-matched object; the second task is used to indicate the probability prediction of the to-be-matched user performing the second operation on the to-be-matched object; based on the common domain data samples and the private domain data samples, train the second sub-network corresponding to the first task and the second sub-network corresponding to the second task respectively.

[0174] In an optional implementation, the target network model also includes a first feature extraction network, a second feature extraction network and a feature fusion network;

[0175] The training module 804 is also used to: perform feature extraction on the common domain data sample through the first feature extraction network to obtain first sample feature data; perform feature extraction on the private domain data sample through the second feature extraction network to obtain second sample feature data; fuse the first sample feature data and the second sample feature data through the feature fusion network to obtain fused sample feature data; perform supervised training on the first network model based on the first sample feature data; and perform supervised training on the second network model based on the fused sample feature data.

[0176] In an optional implementation, the training module 804 is further used to: perform supervised training on the second network model according to the fused sample feature data in combination with the first prediction result of the first network model.

[0177] The model training device of the embodiment of the present application is used to implement the corresponding model training method in the aforementioned method embodiment, and has the beneficial effects of the corresponding method embodiment, which will not be repeated here. In addition, the functional implementation of each module in the model training device of the embodiment of the present application can refer to the description of the corresponding part in the aforementioned method embodiment, which will not be repeated here.

[0178] Reference Fig. 9 , Fig. 9 : is a structural block diagram of a recommendation device according to an embodiment of the present application. The recommendation device provided by the embodiment of the present application includes:

[0179] The information acquisition module 902 is used to acquire the attribute information of the target individual's to-be-recommended object in response to an input operation on the user terminal of the target individual.

[0180] The sending module 904 is used to send the attribute information of the object to be recommended to the target server; the target server is used to process the acquired attribute information of the user to be recommended and the attribute information of the object to be recommended by the data processing method as described in the first aspect, and obtain the matching result of the object to be recommended and the user to be recommended.

[0181] The receiving module 906 is used to receive the matching results fed back by the target server and display them.

[0182] In an optional implementation, the device of this embodiment also includes a query module, which is used to: in response to a first query operation on the user terminal of the target individual, obtain the click data and / or conversion data of the object to be recommended, and the matching results of the object to be recommended fed back by the target server; analyze the matching results of the object to be recommended and the click data and / or conversion data, obtain the analysis results, and display them.

[0183] In an optional implementation, the query module is also used to: in response to a second query operation on the user terminal of the target individual, obtain a reference matching result for the object to be recommended fed back by the target server, the reference matching result being obtained by the target server through processing the attribute information of the user to be recommended and the attribute information of the object to be recommended by a preset network model; analyze the reference matching result and the click data and / or conversion data to obtain a reference analysis result; and display the analysis result and the reference analysis result.

[0184] The recommendation device of the embodiment of the present application is used to implement the corresponding recommendation method in the aforementioned method embodiment, and has the beneficial effects of the corresponding method embodiment, which will not be repeated here. In addition, the functional implementation of each module in the recommendation device of the embodiment of the present application can refer to the description of the corresponding part in the aforementioned method embodiment, which will not be repeated here.

[0185] Reference Fig.10 , shows a schematic diagram of the structure of an electronic device according to an embodiment of the present application. The specific embodiment of the present application does not limit the specific implementation of the electronic device.

[0186] like Fig.10 As shown, the electronic device may include: a processor (processor) 1002 , a communication interface (Communications Interface) 1004 , a memory (memory) 1006 , and a communication bus 1008 .

[0187] in:

[0188] The processor 1002 , the communication interface 1004 , and the memory 1006 communicate with each other via a communication bus 1008 .

[0189] The communication interface 1004 is used to communicate with other electronic devices or servers.

[0190] The processor 1002 is used to execute the program 1010, and specifically can execute the relevant steps in the above method embodiment.

[0191] Specifically, the program 1010 may include program codes, which include computer operation instructions.

[0192] The processor 1002 may be a CPU, or an application specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present application. The one or more processors included in the smart device may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0193] The memory 1006 is used to store the program 1010. The memory 1006 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0194] The program 1010 may be specifically used to enable the processor 1002 to execute the relevant steps in the above method embodiment.

[0195] The specific implementation of each step in program 1010 can refer to the corresponding description of the corresponding step in the above method embodiment, which will not be repeated here. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the above-described devices and modules can refer to the corresponding process description in the above method embodiment, which will not be repeated here.

[0196] The present application also provides a computer storage medium on which a computer program is stored, and when the program is executed by a processor, the method described in any of the above-mentioned multiple method embodiments is implemented. The computer storage medium includes but is not limited to: a compact disc read-only memory (CD-ROM), a random access memory (RAM), a floppy disk, a hard disk or a magneto-optical disk, etc.

[0197] An embodiment of the present application also provides a computer program product, including computer instructions, which instruct a computing device to execute operations corresponding to the method of any one of the above-mentioned multiple method embodiments.

[0198] In addition, it should be noted that the user-related information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to sample data used to train the model, data used for analysis, stored data, displayed data, etc.) involved in the embodiments of the present application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of the relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0199] It should be pointed out that, according to the needs of implementation, the various components / steps described in the embodiments of the present application can be split into more components / steps, or two or more components / steps or partial operations of components / steps can be combined into new components / steps to achieve the purpose of the embodiments of the present application.

[0200] The above-mentioned method according to the embodiment of the present application can be implemented in hardware, firmware, or implemented as software or computer code that can be stored in a recording medium (such as a CD ROM, RAM, floppy disk, hard disk or magneto-optical disk), or implemented as a computer code originally stored in a remote recording medium or a non-temporary machine-readable medium downloaded through a network and stored in a local recording medium, so that the method described herein can be stored in such software processing on a recording medium using a general-purpose computer, a special-purpose processor or programmable or special-purpose hardware (such as an ASIC or FPGA). It can be understood that a computer, a processor, a microprocessor controller or programmable hardware includes a storage component (e.g., RAM, ROM, flash memory, etc.) that can store or receive software or computer code, and when the software or computer code is accessed and executed by a computer, a processor or hardware, the method described herein is implemented. In addition, when a general-purpose computer accesses a code for implementing the method or method shown here, the execution of the code converts the general-purpose computer into a special-purpose computer for executing the method shown here.

[0201] Those of ordinary skill in the art will appreciate that the units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the embodiments of the present application.

[0202] The above implementation methods are only used to illustrate the embodiments of the present application, and are not limitations on the embodiments of the present application. Ordinary technicians in the relevant technical field can make various changes and modifications without departing from the spirit and scope of the embodiments of the present application. Therefore, all equivalent technical solutions also belong to the scope of the embodiments of the present application. The scope of patent protection of the embodiments of the present application should be limited by the claims.

Claims

1. A data processing method, comprising: Acquire input data, wherein the input data includes attribute information of a user to be matched and attribute information of an object to be matched; Inputting the input data into a target network model for predicting a matching target object for a user, the target network model comprising a first network model and a second network model, the first network model being a network model trained based on common domain data of a target group, and the second network model being a network model trained based on the common domain data and private domain data of some individuals in the target group; The input data is processed by the first network model or the second network model to obtain a matching result between the to-be-matched user and the to-be-matched object.

2. The method according to claim 1, wherein: The target group includes a first group and a second group, and the part of individuals belongs to the second group; The step of processing the input data by using the first network model or the second network model to obtain a matching result between the to-be-matched user and the to-be-matched object includes: If the attribute information of the to-be-matched object indicates that the to-be-matched object corresponds to the first group, processing the input data through the first network model to obtain a matching result for the to-be-matched user and the to-be-matched object; If the attribute information of the object to be matched indicates that the object to be matched corresponds to the second group, the input data is processed by the second network model to obtain a matching result for the user to be matched and the object to be matched.

3. The method according to claim 2, wherein: The step of processing the input data by using the first network model to obtain a matching result between the to-be-matched user and the to-be-matched object includes: Based on the input data, probability predictions are respectively performed to indicate that the user to be matched performs a first operation and a second operation on the object to be matched, and corresponding first operation prediction results and second operation prediction results are obtained; based on the first operation prediction result and the second operation prediction result, a matching result for the user to be matched and the object to be matched is obtained.

4. The method according to claim 2 or 3, wherein: The step of processing the input data by using the second network model to obtain a matching result between the to-be-matched user and the to-be-matched object includes: Based on the input data, probability predictions are respectively performed to indicate that the user to be matched performs a first operation and a second operation on the object to be matched, and corresponding third operation prediction results and fourth operation prediction results are obtained; based on the third operation prediction result and the fourth operation prediction result, a matching result for the user to be matched and the object to be matched is obtained.

5. The method according to claim 1, wherein: The obtaining of input data comprises: Obtaining initial attribute information of the object to be matched and the target individual to which the object to be matched belongs; According to the preset association relationship between target individuals and / or the association relationship between objects possessed by the target individuals, the initial attribute information of the object to be matched is replaced with the attribute information of the object to be matched; The input data is determined according to the attribute information of the object to be matched and the obtained attribute information of the user to be matched.

6. The method according to claim 1, wherein: The step of processing the input data by using the first network model or the second network model to obtain a matching result between the to-be-matched user and the to-be-matched object includes: Processing the input data by using the first network model or the second network model to obtain initial prediction results for the to-be-matched user and the to-be-matched object; The initial prediction result is weighted and adjusted to obtain a matching result for the to-be-matched user and the to-be-matched object.

7. A model training method, comprising: Acquire training data for training a target network model, wherein the training data includes common domain data samples of a target group and private domain data samples of some individuals in the target group, wherein both the common domain data samples and the private domain data samples include attribute information of sample users and attribute information of sample objects to be matched; Based on the common domain data samples, a first network model of the target network model is trained, and based on the common domain data samples and the private domain data samples, a second network model of the target network model is trained.

8. The method according to claim 7, wherein: The first network model includes a first subnetwork corresponding to the first task and a first subnetwork corresponding to the second task, and the second network model includes a second subnetwork corresponding to the first task and a second subnetwork corresponding to the second task; Based on the common domain data samples, a first network model of the target network model is trained, and based on the common domain data samples and the private domain data samples, a second network model of the target network model is trained, including: Based on the common domain data samples, the first sub-network corresponding to the first task and the first sub-network corresponding to the second task are trained respectively; the first task is used to indicate the probability prediction of the to-be-matched user performing a first operation on the to-be-matched object; the second task is used to indicate the probability prediction of the to-be-matched user performing a second operation on the to-be-matched object; Based on the common domain data samples and the private domain data samples, the second sub-network corresponding to the first task is trained, and the second sub-network corresponding to the second task is trained, respectively.

9. The method according to claim 7 or 8, wherein: The target network model also includes a first feature extraction network, a second feature extraction network and a feature fusion network; The training of the first network model of the target network model based on the common domain data sample, and the training of the second network model of the target network model based on the common domain data sample and the private domain data sample, includes: Extracting features from the common domain data sample through the first feature extraction network to obtain first sample feature data; extracting features from the private domain data sample through the second feature extraction network to obtain second sample feature data; fusing the first sample feature data and the second sample feature data through the feature fusion network to obtain fused sample feature data; Based on the first sample feature data, supervised training is performed on the first network model; based on the fused sample feature data, supervised training is performed on the second network model.

10. The method according to claim 9, wherein: The training of the second network model based on the fused sample feature data includes: performing supervised training on the second network model according to the fused sample feature data in combination with a first prediction result of the first network model.

11. A recommendation method, applied to a user terminal of a target individual, comprising: In response to an input operation on a user terminal of the target individual, acquiring attribute information of an object to be recommended of the target individual; Sending the attribute information of the object to be recommended to the target server; The target server is used to process the acquired attribute information of the user to be recommended and the attribute information of the object to be recommended by the data processing method according to any one of claims 1 to 6, so as to obtain a matching result between the object to be recommended and the user to be recommended; Receive the matching result fed back by the target server and display it.

12. The method according to claim 11, wherein: After receiving the matching result fed back by the target server, the method further includes: In response to a first query operation on a user terminal of the target individual, acquiring click data and / or conversion data of the object to be recommended, and a matching result of the object to be recommended fed back by the target server; The matching result of the object to be recommended and the first operation data and / or the second operation data are analyzed to obtain an analysis result, and the analysis result is displayed.

13. The method according to claim 12, wherein: The method further comprises: In response to a second query operation on the user terminal of the target individual, obtaining a reference matching result for the object to be recommended fed back by the target server, wherein the reference matching result is obtained by the target server processing the attribute information of the user to be recommended and the attribute information of the object to be recommended through a preset network model; Analyze the reference matching result and the first operation data and / or the second operation data to obtain a reference analysis result; The analysis results and the reference analysis results are displayed.

14. An electronic device comprising: A processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform an operation corresponding to the method as described in any one of claims 1-6, or to perform an operation corresponding to the method as described in any one of claims 7-10, or to perform an operation corresponding to the method as described in claims 11-13.

Images