Resource recall method, device and storage medium

By splicing and processing the user and resource characteristics and matching them with deep learning models, the problem of low advertising screening efficiency is solved, and the target resources are quickly screened, and the response speed is improved.

CN114581115BActive Publication Date: 2024-12-03BEIJING DAJIA INTERNET INFORMATION TECH CO LTD
View PDF 7 Cites 0 Cited by

Patent Information

Application Number
CN202011375002.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-11-30
Publication Date
2024-12-03
Estimated Expiration
2040-11-30

AI Technical Summary

Technical Problem

When existing advertising screening technology processes a large number of advertising units and user information, the matching efficiency is low, resulting in a long time-consuming screening process and affecting the response speed.

Method used

By obtaining the feature information of the user set and resource set, performing splicing processing, generating splicing user features and splicing resource features, using deep learning models for matching processing, and generating predicted matching data, thereby filtering out target resources and reducing the parameters and times of matching calculations. .

Benefits of technology

It improves the response speed of resource recall, improves the efficiency of advertising screening, and solves the problem of low resource screening efficiency in the existing technology.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114581115B_ABST
    Figure CN114581115B_ABST
Patent Text Reader

Abstract

The present disclosure relates to a resource recall method, device and storage medium. The method obtains user information of multiple users in a user set and resource information of multiple resources in a resource set, and splices the user features of all user information in the user set and the resource features of all resource information in the resource set to obtain spliced ​​user features and spliced ​​resource features, and then performs matching processing on the spliced ​​user features and the spliced ​​resource features to obtain predicted matching data. According to the predicted matching data, the target resources to be recalled by the users corresponding to each user information are screened out from the resource set. The predicted matching data obtained by matching the spliced ​​user features and the spliced ​​resource features characterizes the matching relationship between all users in the user set and all resources in the resource set, and achieves the matching results between multiple user information and multiple resource information obtained through one matching process, which can improve the response speed of resource recall.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a resource recall method, device and storage medium. Background Art

[0002] Advertisements are a special type of information and content resource. Compared to general information and content, advertisements have higher requirements for user targeting due to their relatively high display costs. Usually, each ad unit has its target audience, and advertisers always hope that the ad unit can be accurately displayed to the target audience it is designed for. To this end, the advertising platform will set audience display rules for each ad unit. When a user's client requests to display an ad in a certain advertising booth, the advertising platform needs to select the appropriate ad from the alternative ads.

[0003] Usually, this screening process is to match each ad unit with each user, determine whether the user meets the target demographic characteristics of the ad unit, and then decide which ads to serve to the user. When the number of candidate ads is very large, for example, up to 10,000, the number of parameters involved in the screening process is huge, and each calculation can only obtain the matching probability between an ad unit and a user. After multiple calculations to obtain the matching probability between the user and each ad unit, further ad screening can be performed, resulting in low screening efficiency. Summary of the invention

[0004] The present disclosure provides a resource recall method, device and storage medium, which can improve resource screening efficiency. The technical solution of the present disclosure is as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, a resource recall method is provided, including:

[0006] Acquire a user set and a resource set, wherein the user set includes user information of multiple users, and the resource set includes resource information of multiple resources to be delivered;

[0007] splicing user features of all user information in the user set to obtain spliced ​​user features, and splicing resource features of all resource information in the resource set to obtain spliced ​​resource features;

[0008] Performing matching processing on the spliced ​​user features and the spliced ​​resource features to obtain predicted matching data, wherein the predicted matching data includes user matching data corresponding to each user information in the user set and / or resource matching data corresponding to each resource information in the resource set, wherein the user matching data is used to represent the matching probability between the corresponding user information and each resource information in the resource set, and the resource matching data is used to represent the matching probability between the corresponding resource information and each user information in the user set;

[0009] According to the predicted matching data, target resources to be recalled by users corresponding to each user information are screened out from the resource set.

[0010] In an exemplary embodiment, the method further includes: performing feature extraction on the user information and the resource information respectively to obtain user features of each user information in the user set and resource features of each resource information in the resource set.

[0011] In an exemplary embodiment, the step of selecting target resources to be recalled by users corresponding to each user information from the resource set according to the predicted matching data includes:

[0012] Determine a target matching probability from a plurality of matching probabilities included in each user matching data or a plurality of matching probabilities included in each resource matching data;

[0013] The resource information and user information associated with the target matching probability are determined, and the resource information associated with the target matching probability is used as the target resource to be recalled by the user corresponding to the user information associated with the target matching probability.

[0014] In an exemplary embodiment, determining the target matching probability from a plurality of matching probabilities included in each user matching data or a plurality of matching probabilities included in each resource matching data includes:

[0015] The matching probabilities included in each user matching data are sorted, and the target matching probability is determined based on the sorting result; or, the matching probabilities included in each resource matching data are sorted, and the target matching probability is determined based on the sorting result; or, the target matching probability is determined based on the comparison result between the matching probabilities included in each user matching data and a preset threshold; or, the target matching probability is determined based on the comparison result between the matching probabilities included in each resource matching data and a preset threshold.

[0016] In an exemplary embodiment, determining resource information and user information associated with the target matching probability includes:

[0017] Obtaining a resource identifier and a user identifier corresponding to the target matching probability;

[0018] determining resource information associated with the target matching probability according to the resource identifier, and determining user information associated with the target matching probability according to the user identifier,

[0019] Among them, the user identifier is the input sequence information or preset user number information of the user characteristics of the user information input into the prediction model, and the resource identifier is the input sequence information or preset resource number information of the resource characteristics of the resource information input into the prediction model.

[0020] In an exemplary embodiment, the user information in the user set and the resource information in the resource set are input into a prediction model, and the prediction matching data is obtained through the prediction model, and the prediction model is trained by the following method:

[0021] Acquire sample data, the sample data including a sample user set, a sample resource set and a label set, the sample user set including multiple sample user information, the sample resource set including multiple sample resource information, the label set including a label value corresponding to each sample user information or each sample resource information, each label value being used to indicate whether a piece of sample user information matches a piece of sample resource information;

[0022] Inputting the sample user information and the sample resource information into a deep learning model, obtaining sample splicing user features and sample splicing resource features, performing matching processing on the sample splicing user features and the sample splicing resource features, and obtaining sample prediction matching data;

[0023] Determining label data corresponding to the sample predicted matching data according to the sample predicted matching data and the label value in the sample data;

[0024] Calculate the loss value based on the sample prediction matching data and the label data;

[0025] The deep learning model is trained according to the loss value to obtain a prediction model.

[0026] In an exemplary embodiment, the acquiring of sample splicing user features and sample splicing resource features, matching the sample splicing user features and the sample splicing resource features, and obtaining sample prediction matching data includes:

[0027] splicing sample user features of all sample user information in the sample user set to obtain sample splicing user features, and splicing sample resource features of all sample resource information in the sample resource set to obtain sample splicing resource features;

[0028] The sample splicing user features and the sample splicing resource features are matched to obtain sample prediction matching data, wherein the sample prediction matching data includes sample user matching data corresponding to each sample user information in the sample user set and / or sample resource matching data corresponding to each sample resource information in the sample resource set, wherein the sample user matching data is used to characterize the matching probability between the corresponding sample user information and each sample resource information in the sample resource set, and the sample resource matching data is used to characterize the matching probability between the corresponding sample resource information and each sample user information in the sample user set.

[0029] In an exemplary embodiment, the method further includes: performing feature extraction on the sample user information to obtain sample user features of the sample user information; performing feature extraction on the sample resource information to obtain sample resource features of the sample resource information.

[0030] In an exemplary embodiment, determining the label data corresponding to the sample predicted matching data according to the sample predicted matching data and the label value in the sample data includes:

[0031] Determine a sample user identifier and a sample resource identifier associated with each matching probability of each sample user matching data in the sample matching data;

[0032] Acquire a label value corresponding to each matching probability according to the sample user identifier and the sample resource identifier associated with the matching probability;

[0033] Generate label data according to the label value corresponding to each matching probability and the ranking of each matching probability in the sample matching data, wherein the ranking of each label value in the label data is the same as the ranking of the matching probability corresponding to each label value in the sample matching data.

[0034] In an exemplary embodiment, the method also includes: inputting the sample user information into a user feature extraction network of a deep learning model, and using the user feature extraction network to perform feature extraction on the sample user information to obtain sample user features of the sample user information; inputting the sample resource information into a resource feature extraction network of a deep learning model, and using the resource feature extraction network to perform feature extraction on the sample resource information to obtain sample resource features of the sample resource information.

[0035] According to a second aspect of an embodiment of the present disclosure, a resource recall device is provided, including:

[0036] An information acquisition unit is configured to acquire a user set and a resource set, wherein the user set includes user information of a plurality of users, and the resource set includes resource information of a plurality of resources to be delivered;

[0037] The predicted matching data acquisition unit is configured to splice user features of all user information in the user set to obtain spliced ​​user features, and splice resource features of all resource information in the resource set to obtain spliced ​​resource features; match the spliced ​​user features and the spliced ​​resource features to obtain predicted matching data, wherein the predicted matching data includes user matching data corresponding to each user information in the user set and / or resource matching data corresponding to each resource information in the resource set, wherein the user matching data is used to represent the matching probability between the corresponding user information and each resource information in the resource set, and the resource matching data is used to represent the matching probability between the corresponding resource information and each user information in the user set;

[0038] The target resource determination unit is configured to filter out target resources that need to be recalled by users corresponding to each user information from the resource set according to the predicted matching data.

[0039] In an exemplary embodiment, the predicted matching data acquisition unit is configured to perform feature extraction on the user information and the resource information respectively to obtain user features of each user information in the user set and resource features of each resource information in the resource set.

[0040] In an exemplary embodiment, the target resource determination unit includes:

[0041] A target matching probability selection module is configured to determine a target matching probability from a plurality of matching probabilities included in each user matching data or a plurality of matching probabilities included in each resource matching data;

[0042] The target resource determination module is configured to determine resource information and user information associated with the target matching probability, and use the resource information associated with the target matching probability as the target resource to be recalled by the user corresponding to the user information associated with the target matching probability.

[0043] In an exemplary embodiment, the target matching probability selection module is configured to: sort the matching probabilities contained in each user matching data, and determine the target matching probability based on the sorting result; or sort the matching probabilities contained in each resource matching data, and determine the target matching probability based on the sorting result; or determine the target matching probability based on the comparison result between the matching probabilities contained in each user matching data and a preset threshold; or determine the target matching probability based on the comparison result between the matching probabilities contained in each resource matching data and a preset threshold.

[0044] In an exemplary embodiment, the target resource determination module includes:

[0045] An identification acquisition submodule, configured to acquire a resource identification and a user identification corresponding to the target matching probability;

[0046] A determination submodule is configured to determine the resource information associated with the target matching probability based on the resource identifier, and to determine the user information associated with the target matching probability based on the user identifier, wherein the user identifier is the input sequence information or preset user number information of the user characteristics of the user information input into the prediction model, and the resource identifier is the input sequence information or preset resource number information of the resource characteristics of the resource information input into the prediction model.

[0047] In an exemplary embodiment, the device further includes a model acquisition unit, which is used to process the user information in the user set and the resource information in the resource set to obtain the predicted matching data; the model acquisition unit includes:

[0048] a sample acquisition module configured to acquire sample data, wherein the sample data includes a sample user set, a sample resource set, and a label set, wherein the sample user set includes a plurality of sample user information, the sample resource set includes a plurality of sample resource information, and the label set includes a label value corresponding to each sample user information or each sample resource information, wherein each label value is used to indicate whether a piece of sample user information matches a piece of sample resource information;

[0049] The model training module is configured to input the sample user information and the sample resource information into a deep learning model, obtain sample splicing user features and sample splicing resource features, match the sample splicing user features and the sample splicing resource features, and obtain sample prediction matching data; determine the label data corresponding to the sample prediction matching data according to the sample prediction matching data and the label value in the sample data; calculate the loss value according to the sample prediction matching data and the label data; and train the deep learning model according to the loss value to obtain a prediction model.

[0050] In an exemplary embodiment, the model training module is configured to:

[0051] splicing sample user features of all sample user information in the sample user set to obtain sample splicing user features, and splicing sample resource features of all sample resource information in the sample resource set to obtain sample splicing resource features;

[0052] The sample splicing user features and the sample splicing resource features are matched to obtain sample prediction matching data, wherein the sample prediction matching data includes sample user matching data corresponding to each sample user information in the sample user set and / or sample resource matching data corresponding to each sample resource information in the sample resource set, wherein the sample user matching data is used to characterize the matching probability between the corresponding sample user information and each sample resource information in the sample resource set, and the sample resource matching data is used to characterize the matching probability between the corresponding sample resource information and each sample user information in the sample user set.

[0053] In an exemplary embodiment, the model training module is configured to perform feature extraction on the sample user information to obtain sample user features of the sample user information; and perform feature extraction on the sample resource information to obtain sample resource features of the sample resource information.

[0054] In an exemplary embodiment, the model training module includes:

[0055] a sample identification determination submodule, configured to determine a sample user identification and a sample resource identification associated with each matching probability of each sample user matching data in the sample matching data;

[0056] A label value acquisition submodule is configured to acquire a label value corresponding to each matching probability according to a sample user identifier and a sample resource identifier associated with the matching probability;

[0057] The label data generation submodule is configured to generate label data according to the label value corresponding to each matching probability and the order of each matching probability in the sample matching data, wherein the order of each label value in the label data is the same as the order of the matching probability corresponding to each label value in the sample matching data.

[0058] In an exemplary embodiment, the model training module further includes:

[0059] The sample feature acquisition submodule is configured to input the sample user information into the user feature extraction network of the deep learning model, use the user feature extraction network to extract features from the sample user information, and obtain sample user features of the sample user information; input the sample resource information into the resource feature extraction network of the deep learning model, use the resource feature extraction network to extract features from the sample resource information, and obtain sample resource features of the sample resource information.

[0060] According to a third aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0061] processor;

[0062] a memory for storing instructions executable by the processor;

[0063] The processor is configured to execute the instructions to implement the resource recall method as described in the first aspect above.

[0064] According to a fourth aspect of an embodiment of the present disclosure, a storage medium is provided. When instructions in the storage medium are executed by a processor of an electronic device, the electronic device can execute the resource recall method as described in the first aspect.

[0065] According to a fifth aspect of an embodiment of the present disclosure, a computer program product is provided. When instructions in the computer program product are executed by a processor of a computer device, the computer device is enabled to execute the resource recall method described in the first aspect.

[0066] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects:

[0067] The resource recall method, device and storage medium provided by the embodiments of the present disclosure obtain user information of multiple users in a user set and resource information of multiple resources in a resource set, determine user features of each user information and resource features of each resource information, splice user features of all user information in the user set to obtain spliced ​​user features, splice resource features of all resource information in the resource set to obtain spliced ​​resource features, then match the spliced ​​user features with the spliced ​​resource features to obtain predicted matching data, and screen out target resources to be recalled by users corresponding to each user information from the resource set based on the predicted matching data. In the disclosed embodiment, the predicted matching data obtained by matching the spliced ​​user features and the spliced ​​resource features does not represent the matching relationship between a single user information and each resource information, but represents the matching relationship between all users in the user set and all resources in the resource set, specifically including user matching data corresponding to each user information in the user set and / or resource matching data corresponding to each resource information in the resource set, wherein the user matching data is used to represent the matching probability between the corresponding user information and each resource information in the resource set, and the resource matching data is used to represent the matching probability between the corresponding resource information and each user information in the user set; the matching results between multiple user information and multiple resource information are obtained through one matching process, the number of parameters involved in the calculation in the matching process between user information and resource information is reduced, and the number of matching calculations is reduced, thereby improving the response speed of resource recall and solving the problem of low resource screening efficiency in related technologies.

[0068] In the disclosed embodiment, sample user information in a sample user set and sample resource information in a sample resource set are input into a deep learning model, and label data corresponding to the sample prediction data is determined based on the sample prediction matching data output by the deep learning model in combination with the label value in the sample data. The deep learning model is then trained based on the loss value between the label data and the sample prediction matching data to obtain a prediction model, which can enable the prediction model to automatically learn the ability to output the matching probabilities between multiple user information and multiple resource information in the form of overall data, thereby improving the data processing efficiency of the model without reducing the prediction ability of the model.

[0069] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0071] Figure 1The diagram is a schematic diagram of an implementation environment of a resource recall method according to an exemplary embodiment.

[0072] Figure 2 It is a schematic diagram of the network structure of a model used to predict click rate in related technologies.

[0073] Figure 3 The figure is a flowchart of a resource recall method according to an exemplary embodiment.

[0074] Figure 4 The figure is a flow chart of a resource recall method according to an exemplary embodiment.

[0075] Figure 5 The figure is a flowchart of a method for training a prediction model according to an exemplary embodiment.

[0076] Figure 6 The figure is a schematic diagram of a network structure of a prediction model according to an exemplary embodiment.

[0077] Figure 7 The figure is a flowchart of a method for recalling resources based on a prediction model according to an exemplary embodiment.

[0078] Figure 8 The figure is a schematic diagram of the structure of a resource recall device according to an exemplary embodiment.

[0079] Fig. 9 It is a structural diagram of another resource recall device according to an exemplary embodiment.

[0080] Fig.10 The invention is a block diagram of a server showing a resource recall method according to an exemplary embodiment. DETAILED DESCRIPTION

[0081] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0082] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0083] Figure 1 is a schematic diagram of an implementation environment of a resource recall method according to an exemplary embodiment; please refer to Figure 1 , the implementation environment includes: client 01 and server 03.

[0084] The client 01 may include physical devices such as smart phones, tablet computers, laptop computers, digital assistants, smart wearable devices, and vehicle-mounted terminals, and may also include software running in physical devices. The client 01 may communicate with the server 03 based on a browser / server mode (B / S) or a client / server mode (C / S).

[0085] The server 03 may include an independently operated server, or a distributed server, or a server cluster consisting of multiple servers.

[0086] The server 03 is used to receive a page access request submitted by a user through a client, and obtain the user information of the user and the resource information of the resources to be delivered, wherein the resources to be delivered can be submitted to the server 03 by the resource owner through the resource delivery client. After receiving the resources to be delivered, the server can store the resources to be delivered in a resource pool. When the server 03 receives the page access request of the user, the resources to be delivered in the resource pool can be matched with the user information, and at least one resource with a high degree of matching with the user information can be found and recalled, and then the recalled resources can be delivered to the user. In the disclosed embodiment, the server 03 extracts features from multiple user information and multiple resource information respectively, splices the extracted multiple user features to obtain spliced ​​user features, splices the extracted multiple resource features to obtain spliced ​​resource features, and then matches the spliced ​​user features with the spliced ​​resource features to obtain predicted matching data, and then selects the target resources that each user needs to recall from multiple resources to be delivered according to the predicted matching data. The prediction matching data includes user matching data corresponding to each user information in the user set and / or resource matching data corresponding to each resource information in the resource set, wherein the user matching data is used to characterize the matching probability between the corresponding user information and each resource information in the resource set, and the resource matching data is used to characterize the matching probability between the corresponding resource information and each user information in the user set. By splicing user features and resource features separately during the prediction process, the parameters of model processing and output are reduced, thereby shortening the prediction time, improving the response speed to user requests, and optimizing the user experience.

[0087] When a user opens an application on the client side or browses the content in the application, the client sends a resource recall request to the server. The server can obtain the user information and the resource information of the resources to be delivered, and input the user information and resource information into the trained model to predict the click-through rate. Among them, the user information includes but is not limited to the user's gender, age, etc., the resources to be delivered can be advertising resources, and the resource information includes but is not limited to the logo, aspect ratio, video length, etc. of the advertisement. The click-through rate prediction is to predict the probability of a user clicking on the advertisement to be delivered or the probability of a user clicking on the advertisement to be delivered. In the related art, the model used for click-through rate prediction is generally obtained by training a multi-layer perceptron neural network, and its network structure is as follows: Figure 2 As shown, it includes a feature layer, a hidden layer, an output layer and a prediction layer. The network structure is used to match user information and advertising information. It includes the process of processing user information and advertising information separately and matching. Therefore, the feature layer includes a user feature layer and an advertising feature layer, the hidden layer includes a user hidden layer and an advertising hidden layer, and the output layer includes a user output layer and an advertising output layer. After the user information is input into the network, the user feature layer extracts the features of the user information, and the user hidden layer processes and outputs the features from the user output layer. After the advertising information is input into the network, the advertising feature layer extracts the features of the advertising information, and the advertising hidden layer processes and outputs the features from the advertising output layer. The prediction layer matches the output of the user output layer with the output of the advertising output layer, and outputs the matching probability of the user and the advertisement. The matching probability is the click-through rate of the user clicking the advertisement. In the process of using the network structure to predict the click-through rate, since the user information of a single user and the advertising information of a single advertisement are used as input for matching prediction, when the number of advertisement recall requests at the same time is large, the feedback of the advertisement recall result for a single user will inevitably be delayed, and the speed response becomes the bottleneck of the online advertisement recall request. For example, when the number of user information is U, the number of network parameters is: U×H1+H1×OU+I×H2+H2×OI, where U×H1+H1×OU is the number of user-side parameters, and I×H2+H2×OI is the number of advertising-side parameters.

[0088] In order to solve the technical problem in the related art that the model has a slow response speed to multiple advertising resource recall requests, the embodiments of the present disclosure propose to optimize the network structure of the model, change the design of the model that matches a single user with a single advertising resource, use multiple users to match multiple advertising resources, and achieve the prediction probability of multiple users clicking on multiple advertising resources through a single input, reduce the number of parameters that the model needs to process, and improve the response speed to advertising resource recall requests.

[0089] According to the above inventive ideas, the embodiments of the present disclosure provide Figure 3 A flowchart of a resource recall method is shown in FIG. Figure 3A resource recall method disclosed in the present invention comprises the following steps:

[0090] S301, obtaining a user set and a resource set, wherein the user set includes user information of multiple users, and the resource set includes resource information of multiple resources to be delivered.

[0091] S303, splicing user features of all user information in the user set to obtain spliced ​​user features, and splicing resource features of all resource information in the resource set to obtain spliced ​​resource features.

[0092] S305, matching processing is performed on the spliced ​​user features and the spliced ​​resource features to obtain predicted matching data, where the predicted matching data includes user matching data corresponding to each user information in the user set and / or resource matching data corresponding to each resource information in the resource set.

[0093] The user matching data is used to represent the matching probability between the corresponding user information and each resource information in the resource set, and the resource matching data is used to represent the matching probability between the corresponding resource information and each user information in the user set.

[0094] S307: Filter out target resources that need to be recalled by users corresponding to each user information from the resource set according to the predicted matching data.

[0095] The resource recall method provided by the embodiment of the present disclosure obtains user information of multiple users in a user set and resource information of multiple resources in a resource set, determines user features of each user information and resource features of each resource information, splices user features of all user information in the user set to obtain spliced ​​user features, splices resource features of all resource information in the resource set to obtain spliced ​​resource features, then matches the spliced ​​user features with the spliced ​​resource features to obtain predicted matching data, and based on the predicted matching data, screens out target resources to be recalled by users corresponding to each user information from the resource set. In the disclosed embodiment, the predicted matching data obtained by matching the spliced ​​user features and the spliced ​​resource features does not represent the matching relationship between a single user information and each resource information, but represents the matching relationship between all users in the user set and all resources in the resource set, specifically including user matching data corresponding to each user information in the user set and / or resource matching data corresponding to each resource information in the resource set, wherein the user matching data is used to represent the matching probability between the corresponding user information and each resource information in the resource set, and the resource matching data is used to represent the matching probability between the corresponding resource information and each user information in the user set; the matching results between multiple user information and multiple resource information are obtained through one matching process, the number of parameters involved in the calculation in the matching process between user information and resource information is reduced, and the number of matching calculations is reduced, thereby improving the response speed of resource recall and solving the problem of low resource screening efficiency in related technologies.

[0096] In a possible implementation, the data in the user set and the resource set may be processed by a model to obtain predicted matching data. Figure 4 is a flow chart showing a resource recall method according to an exemplary embodiment. Figure 4 The resource recall method disclosed in the present invention may include two parts, namely, a model training part and a model application part. The model training part mainly trains a deep learning model based on training samples (sample user set, sample resource set and label set) to obtain a prediction model; the model application part mainly inputs the user information of the user set and the resource information of the resource set into the prediction model, and outputs the prediction matching data. Then, based on the prediction matching data, the target resources to be recalled by the users corresponding to each user information can be screened out from the resource set.

[0097] Figure 5 FIG. 1 is a flowchart of a method for training a prediction model according to an exemplary embodiment. Figure 5 As shown, the method is used in a server and includes the following steps.

[0098] S501, obtaining sample data, where the sample data includes a sample user set, a sample resource set and a label set, where the sample user set includes multiple sample user information, the sample resource set includes multiple sample resource information, and the label set includes a label value corresponding to each sample user information or each sample resource information, where each label value is used to indicate whether a sample user information matches a sample resource information.

[0099] In a possible implementation, the sample resources may be advertising resources. Specifically, sample data may be generated based on advertisements exposed by the advertising system and the users to whom the advertisements are exposed. N advertisements are delivered to M users, where M and N are integers greater than 1. Statistics are given of clicks by each of the M users on each of the N advertisements. If the user clicks on the advertisement, it is determined that the user information and the advertisement match each other, and the user information and the advertisement clicked by the user may be used as positive samples. If the user does not click on the advertisement, it is determined that the user information and the advertisement do not match each other, and the user and the advertisement not clicked by the user may be used as negative samples. Furthermore, label values ​​corresponding to the positive samples and negative samples may be assigned. The label values ​​are used to indicate the positive and negative sample properties of the samples (that is, whether the user information corresponding to the samples matches the advertisements). Specifically, the label values ​​may be reflected in the click rate of the users on the advertisements. For example, the label value corresponding to the positive sample may be assigned 100%, and the label value corresponding to the negative sample may be assigned 0%. Furthermore, sample data can be generated based on positive samples and corresponding label values, negative samples and corresponding label values. The sample data includes advertising information (i.e., resource information) of N advertisements, user information of M users, and label values ​​corresponding to each of the M users and each of the N advertisements.

[0100] It should be noted that, since the samples are usually some offline data, the sample user information included in the above sample data can be called the offline information of the sample user, which may include but is not limited to gender, age, place of residence, shopping records and consumption records, etc. The above sample advertising information may at least include text information and picture information, wherein the text information may include but is not limited to word vector features, and the picture information may include but is not limited to pixel features, etc.

[0101] In one example, the sample user information and sample advertisement information may be obtained through a feature extraction tool. The sample user information and sample advertisement information obtained may be preprocessed. The preprocessing may include but is not limited to feature filtering, default value filling, and normalization. In addition, the obtained sample user information and sample advertisement information may be stored.

[0102] S503, inputting the sample user information and the sample resource information into the deep learning model, obtaining the sample splicing user features and the sample splicing resource features, performing matching processing on the sample splicing user features and the sample splicing resource features, and obtaining the sample prediction matching data.

[0103] In a possible implementation, after the sample user information and the sample resource information are input into the deep learning model, the deep learning model is used to perform the following processing:

[0104] 1) Extracting features from sample user information to obtain sample user features of the sample user information; extracting features from sample resource information to obtain sample resource features of the sample resource information;

[0105] 2) splicing the sample user features of all sample user information in the sample user set to obtain sample splicing user features, and splicing the sample resource features of all sample resource information in the sample resource set to obtain sample splicing resource features;

[0106] 3) Matching the sample splicing user features and the sample splicing resource features to obtain sample prediction matching data, the sample prediction matching data including sample user matching data corresponding to each sample user information in the sample user set and / or sample resource matching data corresponding to each sample resource information in the sample resource set, wherein the sample user matching data is used to characterize the matching probability between the corresponding sample user information and each sample resource information in the sample resource set, and the sample resource matching data is used to characterize the matching probability between the corresponding sample resource information and each sample user information in the sample user set.

[0107] Figure 6 is a schematic diagram of a network structure of a prediction model according to an exemplary embodiment. Figure 6 , including a feature extraction network, a feature splicing network and a prediction network. The feature extraction network is divided into a user feature extraction network and a resource feature extraction network. The feature splicing network is divided into a user feature splicing network and a resource feature splicing network. The user feature extraction network includes a user feature layer, multiple user hidden layers and multiple user output layers. The user output layer corresponds to the user hidden layer one by one, and the outputs of the multiple user output layers serve as the input of the user output splicing layer together. The resource feature extraction network includes a resource feature layer, multiple resource hidden layers and multiple resource output layers. The resource output layer corresponds to the resource hidden layer one by one, and the outputs of the multiple resource output layers serve as the input of the resource output splicing layer together. The prediction network performs probability prediction according to the output of the user output splicing layer and the output of the resource output splicing layer, and the matching prediction data outputted by the prediction network can represent the predicted probability of each user clicking on each resource.

[0108] In a feasible implementation, the sample user information can be input into the user feature extraction network of the deep learning model, and the user feature extraction network is used to extract features from the sample user information to obtain sample user features of the sample user information; the sample resource information can be input into the resource feature extraction network of the deep learning model, and the resource feature extraction network is used to extract features from the sample resource information to obtain sample resource features of the sample resource information. The sample user features of all sample user information in the sample user set are input into the user feature splicing network of the deep learning model, and the user feature splicing network is used to splice the sample user features of all sample user information in the sample user set to obtain sample splicing user features; the sample resource features of all sample resource information in the sample resource set are input into the resource feature splicing network of the deep learning model, and the resource feature splicing network is used to splice the sample resource features of all sample resource information in the sample resource set to obtain sample splicing resource features. The sample splicing user features and the sample splicing resource features are input into the prediction network of the deep learning model, and the prediction network is used to match the sample splicing user features and the sample splicing resource features to obtain sample prediction matching data.

[0109] Specifically, after inputting multiple sample user information into the feature extraction network, the user feature layer extracts the sample user feature vector corresponding to each sample user information, and passes the extracted sample user feature vector of each sample user as a separate parameter to the corresponding hidden layer, and then calculates the result through the weight and activation function, and passes the result to the next layer until it is finally passed to the output layer. Each output layer outputs an output result corresponding to the sample user information, and the output layer passes the output result to the user output splicing layer. The user output splicing layer splices the output results from each output layer to obtain a feature vector corresponding to multiple sample user information. The resource feature extraction network includes a resource feature layer, a hidden layer and an output layer. When multiple sample resource information is input into the model, the resource feature layer extracts the sample resource feature vector corresponding to each sample resource information, and passes the extracted sample resource feature vector of each sample resource to the corresponding hidden layer as a separate parameter. Then, the result is calculated through weights and excitation functions, and the result is passed to the next layer until it is finally passed to the output layer. Each output layer outputs an output result corresponding to the sample resource information. The output layer passes the output result to the resource output splicing layer. The resource output splicing layer splices the output results from each output layer to obtain a feature vector corresponding to multiple sample resource information; the user output splicing layer and the resource output splicing layer respectively pass the spliced ​​feature vectors to the prediction network. The prediction network matches the spliced ​​user feature vectors and the spliced ​​resource feature vectors, and outputs matching prediction data.

[0110] In view of the related art that inputs single user information and single resource information into the model for prediction processing, the model needs to match the single user information with the single resource information one by one when processing multiple user information and multiple resource information, resulting in a large number of matches and low efficiency. The embodiment of the present disclosure transforms the network structure of the model so that the model can process multiple user information and multiple resource information at the same time, output the matching probability between multiple user information and multiple resource information at one time, and improve data processing efficiency.

[0111] S505: Determine label data corresponding to the sample predicted matching data according to the sample predicted matching data and the label value in the sample data.

[0112] Since the matching prediction data output by the deep learning model is an array composed of multiple prediction probabilities, and the label values ​​provided by the sample data are for a single sample user and a single sample resource, the loss value cannot be calculated directly. Therefore, it is necessary to obtain the label values ​​corresponding to the sample user and sample resource corresponding to each prediction probability, and process the label values ​​into a data structure that is compatible with the matching prediction data.

[0113] In one possible implementation, a sample user identifier and a sample resource identifier associated with each matching probability of each sample user matching data in the sample matching data may be determined; a label value corresponding to each matching probability may be obtained based on the sample user identifier and the sample resource identifier associated with the matching probability; and label data may be generated based on the label value corresponding to each matching probability and the order of each matching probability in the sample matching data, wherein the order of each label value in the label data is the same as the order of the matching probability corresponding to each label value in the sample matching data.

[0114] S507, calculating the loss value based on the sample prediction matching data and the label data.

[0115] S509, training a deep learning model according to the loss value to obtain a prediction model.

[0116] In a possible implementation, training the deep learning model according to the loss value may include: adjusting the parameter value of the deep learning model according to the loss value obtained in step S507, and inputting the sample user features and the sample resource features into the deep learning model after adjusting the parameters, until the training stop condition is reached, and using the trained deep learning model as the prediction model. The training stop condition may be that the loss value is lower than a preset loss threshold, or the number of training times reaches a preset number threshold, or the loss value of the current iteration is not significantly reduced relative to the loss value of the previous iteration.

[0117] It should be noted that the embodiment of the present disclosure only takes the execution subject as an example of a server. In another embodiment, the training method provided by the embodiment of the present disclosure can also be executed by a terminal. The embodiment of the present disclosure does not limit the execution subject.

[0118] In the disclosed embodiment, sample user information in a sample user set and sample resource information in a sample resource set are input into a deep learning model, and label data corresponding to the sample prediction data is determined based on the sample prediction matching data output by the deep learning model in combination with the label value in the sample data. The deep learning model is then trained based on the loss value between the label data and the sample prediction matching data to obtain a prediction model, which can enable the prediction model to automatically learn the ability to output the matching probabilities between multiple user information and multiple resource information in the form of overall data, thereby improving the data processing efficiency of the model without reducing the prediction ability of the model.

[0119] Figure 7 is a flow chart of a method for recalling resources based on a prediction model according to an exemplary embodiment, which is applied to a server, see Figure 7 , including the following steps.

[0120] S701, obtaining a user set and a resource set, wherein the user set includes user information of multiple users, and the resource set includes resource information of multiple resources to be delivered.

[0121] Obtain resource recall requests from at least two users. Taking the resource as advertisement as an example, it can be understood that the user's client application will contain some advertising booths for displaying advertisements. When the user logs in to the client application and browses the content in the application, the client sends a resource recall request to the advertising platform, requesting to obtain matching resources for display in the booth. When the user logs in to the client, there is a corresponding relationship between the user and the client, so it can also be regarded as the resource recall request issued by the user. Generally, the resource recall request includes at least the user information that issues the request and the booth information. The user information may include the user's gender, age, region, interests, etc., and the booth information may be, for example, a booth logo.

[0122] The resources to be released are the resources that the resource owner intends to release to users through the resource platform. Resources generally have corresponding audiences. In order to improve the accuracy of resource release, the resource platform needs to match resources and users based on user information and resource information, that is, predict the probability of a user clicking on a resource after a resource is released to the user. The greater the probability, the greater the possibility that the user will click on the resource after the resource is exposed. Generally, the resources to be released provided by the resource owner will be stored in the resource pool of the resource platform. When a resource recall request is received from the user, the user information of the user can be matched with the resource information of each resource in the resource pool to find the most suitable resource for the user based on the predicted click probability. Among them, the resources can be displayed in the form of video or graphic. For video resources, the resource information includes but is not limited to the aspect ratio of the video screen, the playback time of the video, etc. For graphic resources, the resource information includes but is not limited to the aspect ratio of the image, text information, etc.

[0123] S703, obtaining a prediction model.

[0124] In the disclosed embodiment, the prediction model has been trained and stored in the server, and when resource recall is required, the stored prediction model can be obtained. The prediction model can be obtained by training through steps S501 to S509, or can also be obtained by training in other ways.

[0125] S705, input user information and resource information into the prediction model, splice user features of all user information in the user set to obtain spliced ​​user features, splice resource features of all resource information in the resource set to obtain spliced ​​resource features; match the spliced ​​user features and the spliced ​​resource features to obtain predicted matching data, the predicted matching data includes user matching data corresponding to each user information in the user set and / or resource matching data corresponding to each resource information in the resource set, wherein the user matching data is used to characterize the matching probability between the corresponding user information and each resource information in the resource set, and the resource matching data is used to characterize the matching probability between the corresponding resource information and each user information in the user set.

[0126] Specifically, the prediction model processes the input user information and resource information as follows: extract features from the user information and resource information respectively to obtain user features for each user information in the user set and resource features for each resource information in the resource set; splice the user features of all user information in the user set to obtain spliced ​​user features, and splice the resource features of all resource information in the resource set to obtain spliced ​​resource features; match the spliced ​​user features and the spliced ​​resource features to obtain predicted matching data.

[0127] The prediction model includes a feature extraction network, a feature splicing network and a prediction network, wherein the feature extraction network is further divided into a resource feature extraction network and a user feature extraction network, and the feature splicing network is further divided into a resource feature splicing network and a user feature splicing network. After the user information in the user set and the resource information in the resource set are input into the prediction model, the user feature extraction network extracts features from each user information to obtain the user features of each user information, passes the extracted user features to the user feature splicing network, uses the user feature splicing network to splice the user features of all user information in the user set, obtains the spliced ​​user features, and passes the spliced ​​user features to the prediction network; at the same time, the resource feature extraction network extracts features from the resource information in the resource set to obtain the resource features of each resource information, passes the extracted resource features to the resource feature splicing network, uses the resource feature splicing network to splice the resource features of all resource information in the resource set, obtains the spliced ​​resource features, and passes the spliced ​​resource features to the prediction network; after multiple user information is processed by the user feature extraction network and the user feature splicing network, it is obtained to a spliced ​​user feature vector (i.e., spliced ​​user feature); after the resource information of multiple resources is processed by the resource feature extraction network and the resource feature splicing network, a spliced ​​resource feature vector (i.e., spliced ​​resource feature) is obtained. Although the user information and resource information input into the prediction network are multiple respectively, the prediction network only needs to perform matching prediction on the spliced ​​user feature vector and the spliced ​​resource feature vector. Compared with the solution in the related art that takes a single user information and a single resource information as input and performs matching prediction on a single user and a single resource, the embodiment of the present disclosure greatly reduces the number of parameters that the prediction network needs to process when performing matching prediction on multiple user information and multiple resource information, and can output the matching probability of multiple users and multiple resources at one time, shorten the waiting delay for obtaining the matching probability, and improve the response speed to the resource recall request.

[0128] S707: Filter out target resources that need to be recalled by users corresponding to each user information from the resource set according to the predicted matching data.

[0129] Specifically, the target resources that each user needs to recall can be determined through the following steps.

[0130] Step 1: determine a target matching probability from a plurality of matching probabilities included in each user matching data or a plurality of matching probabilities included in each resource matching data.

[0131] When the predicted matching data is user matching data, the target matching probability is determined from the multiple matching probabilities contained in each user matching data; when the predicted matching data is resource matching data, the target matching probability is determined from the multiple matching probabilities contained in each resource matching data; when the predicted matching data includes user matching data and resource matching data, you can choose to determine the target matching probability from the multiple matching probabilities contained in each user matching data, or determine the target matching probability from the multiple matching probabilities contained in each resource matching data.

[0132] In one possible implementation, the matching probabilities contained in each user matching data may be sorted, and the target matching probability may be determined based on the sorting result. For example, for each user matching data, the matching probabilities in the user matching data may be sorted in descending order, and the top N matching probabilities may be selected as the target matching probability, where N is an integer greater than or equal to 1. Alternatively, the matching probabilities contained in each resource matching data may be sorted, and the target matching probability may be determined based on the sorting result. Similarly, for each resource matching data, the matching probabilities in the resource matching data may be sorted in descending order, and the top N matching probabilities may be selected as the target matching probability, where N is an integer greater than or equal to 1.

[0133] In another possible implementation, the target matching probability can be determined based on the comparison result between each matching probability contained in each user matching data and a preset threshold. Exemplarily, the matching probability greater than the preset threshold in the user matching data can be used as the target matching probability; the matching probability greater than the preset threshold in the user matching data can also be used as the candidate matching probability, and then the candidate matching probabilities are sorted in descending order, and the candidate matching probabilities with a preset number of numbers in the front ranking are selected as the target matching probability. Alternatively, the target matching probability can be determined based on the comparison result between each matching probability contained in each resource matching data and a preset threshold. Similarly, the matching probability greater than the preset threshold in the resource matching data can be used as the target matching probability; the matching probability greater than the preset threshold in the resource matching data can also be used as the candidate matching probability, and then the candidate matching probabilities are sorted in descending order, and the candidate matching probabilities with a preset number of numbers in the front ranking are selected as the target matching probability.

[0134] Step 2: Determine resource information and user information associated with the target matching probability, and use the resource information associated with the target matching probability as the target resource that needs to be recalled by the user corresponding to the user information associated with the target matching probability.

[0135] Wherein, determining the resources to be delivered and the users associated with the target matching probability may include: obtaining the resource identifier and user identifier corresponding to the target matching probability; determining the resource information associated with the target matching probability according to the resource identifier, and determining the user information associated with the target matching probability according to the user identifier, wherein the user identifier is the input sequence information of the user feature input prediction model of the user information or the preset user number information, and the resource identifier is the input sequence information of the resource feature input prediction model of the resource information or the preset resource number information. The user number information may be the user's login account in the application, and the resource number information may be the unique identifier of the resource to be delivered in the resource pool.

[0136] After obtaining the target resources that each user needs to recall, the target resources can be recalled. Furthermore, in response to the user's resource recall request, a mapping relationship can be established between the target resources and the booth information corresponding to the resource recall request, so that resources can be subsequently released according to the mapping relationship.

[0137] In the disclosed embodiment, the matching probability can represent the possibility of a user clicking on resource information. The larger the matching probability, the greater the possibility of a user clicking on resource information. By taking one or more matching probabilities with larger probability values ​​as the target matching probability, the user information and resource information associated with the target matching probability are found, and the resource information is used as the target resource that the user needs to recall corresponding to the user information. This can not only quickly determine the resources that the user needs to recall, but also improve the accuracy of resource recall. In the subsequent process of recommending resource information to users, it is beneficial to increase the probability of users clicking on resource information.

[0138] For example, there are user A and user B, resource 1 and resource 2. According to the relevant technical processing, it is necessary to match the characteristics of user A with the characteristics of resource 1 once, match the characteristics of user A with the characteristics of resource 2 once, match the characteristics of user B with the characteristics of resource 1 once, and match the characteristics of user B with the characteristics of resource 2 once. A total of 4 calculations are required. However, the calculation process of the embodiment of the present disclosure is as follows: the features of user A and user B are concatenated to obtain user feature AB, the features of resource 1 and resource 2 are concatenated to obtain resource feature 12, user feature AB is matched with resource feature 12, and predicted matching data A(1,2)B(1,2) or 1(A,B)2(A,B) is output, wherein A(1,2) or 1(A,B) is regarded as a data unit, data unit A(1,2) indicates the matching probability between user A and resource 1 and resource 2 respectively, and data unit 1(A,B) indicates the matching probability between resource 1 and user A and user B respectively. Then, according to the unique identifiers corresponding to users A, B and resources 1, 2 respectively, the predicted matching data is disassembled to obtain the matching probability of each user and each resource. In this way, it is only necessary to perform a matching operation on user feature AB and resource feature 12 once and process two parameters to obtain 4 matching information, which greatly reduces the number of matching operations and can quickly obtain matching results.

[0139] The resource recall method, device and storage medium provided by the embodiments of the present disclosure obtain user information of multiple users in a user set and resource information of multiple resources in a resource set, determine user features of each user information and resource features of each resource information, splice user features of all user information in the user set to obtain spliced ​​user features, splice resource features of all resource information in the resource set to obtain spliced ​​resource features, then match the spliced ​​user features with the spliced ​​resource features to obtain predicted matching data, and screen out target resources to be recalled by users corresponding to each user information from the resource set based on the predicted matching data. In the disclosed embodiment, the predicted matching data obtained by matching the spliced ​​user features and the spliced ​​resource features does not represent the matching relationship between a single user information and each resource information, but represents the matching relationship between all users in the user set and all resources in the resource set, specifically including user matching data corresponding to each user information in the user set and / or resource matching data corresponding to each resource information in the resource set, wherein the user matching data is used to represent the matching probability between the corresponding user information and each resource information in the resource set, and the resource matching data is used to represent the matching probability between the corresponding resource information and each user information in the user set; the matching results between multiple user information and multiple resource information are obtained through one matching process, the number of parameters involved in the calculation in the matching process between user information and resource information is reduced, and the number of matching calculations is reduced, thereby improving the response speed of resource recall and solving the problem of low resource screening efficiency in related technologies.

[0140] Figure 8 FIG. 1 is a schematic diagram showing a structure of a resource recall device according to an exemplary embodiment. Figure 8 , the resource recall device comprises:

[0141] The information acquisition unit 810 is configured to acquire a user set and a resource set, wherein the user set includes user information of multiple users, and the resource set includes resource information of multiple resources to be delivered.

[0142] The predicted matching data acquisition unit 820 is configured to splice the user features of all user information in the user set to obtain the spliced ​​user features, and to splice the resource features of all resource information in the resource set to obtain the spliced ​​resource features; match the spliced ​​user features and the spliced ​​resource features to obtain the predicted matching data, the predicted matching data including user matching data corresponding to each user information in the user set and / or resource matching data corresponding to each resource information in the resource set, wherein the user matching data is used to characterize the matching probability between the corresponding user information and each resource information in the resource set, and the resource matching data is used to characterize the matching probability between the corresponding resource information and each user information in the user set.

[0143] The target resource determination unit 830 is configured to filter out target resources that need to be recalled by users corresponding to each user information from the resource set according to the predicted matching data.

[0144] In a possible implementation, the predicted matching data acquisition unit 820 is further configured to perform feature extraction on the user information and the resource information respectively, to obtain the user feature of each user information in the user set and the resource feature of each resource information in the resource set.

[0145] In one possible implementation, see Fig. 9 , the target resource determination unit 830 includes:

[0146] The target matching probability selection module 831 is configured to determine the target matching probability from the multiple matching probabilities included in each user matching data or the multiple matching probabilities included in each resource matching data;

[0147] The target resource determination module 832 is configured to determine resource information and user information associated with the target matching probability, and use the resource information associated with the target matching probability as the target resource to be recalled by the user corresponding to the user information associated with the target matching probability.

[0148] In one possible implementation, the target matching probability selection module 831 is configured to: sort the matching probabilities contained in each user matching data, and determine the target matching probability based on the sorting result; or sort the matching probabilities contained in each resource matching data, and determine the target matching probability based on the sorting result; or determine the target matching probability based on the comparison result between the matching probabilities contained in each user matching data and a preset threshold; or determine the target matching probability based on the comparison result between the matching probabilities contained in each resource matching data and a preset threshold.

[0149] In a possible implementation, the target resource determination module 832 further includes:

[0150] The identification acquisition submodule 8321 is configured to acquire the resource identification and user identification corresponding to the target matching probability;

[0151] Determine submodule 8322, which is configured to determine resource information associated with the target matching probability based on the resource identifier, and determine user information associated with the target matching probability based on the user identifier, wherein the user identifier is the input sequence information of the user feature input prediction model of the user information or the preset user number information, and the resource identifier is the input sequence information of the resource feature input prediction model of the resource information or the preset resource number information.

[0152] In a possible implementation, the resource recall device further includes a model acquisition unit 910, which is used to process the user information in the user set and the resource information in the resource set to obtain predicted matching data. The model acquisition unit 910 includes:

[0153] The sample acquisition module 911 is configured to acquire sample data, wherein the sample data includes a sample user set, a sample resource set, and a label set, wherein the sample user set includes a plurality of sample user information, the sample resource set includes a plurality of sample resource information, and the label set includes a label value corresponding to each sample user information or each sample resource information, wherein each label value is used to indicate whether a piece of sample user information matches a piece of sample resource information;

[0154] The model training module 912 is configured to input the sample user information and the sample resource information into a deep learning model, obtain sample splicing user features and sample splicing resource features, match the sample splicing user features and the sample splicing resource features, and obtain sample predicted matching data; determine the label data corresponding to the sample predicted matching data based on the sample predicted matching data and the label value in the sample data; calculate the loss value based on the sample predicted matching data and the label data; and train the deep learning model based on the loss value to obtain a prediction model.

[0155] Further, the model training module 912 is also configured to: perform feature extraction on the sample user information to obtain sample user features of the sample user information; perform feature extraction on the sample resource information to obtain sample resource features of the sample resource information; splice the sample user features of all sample user information in the sample user set to obtain sample spliced ​​user features, and splice the sample resource features of all sample resource information in the sample resource set to obtain sample spliced ​​resource features; perform matching processing on the sample spliced ​​user features and the sample spliced ​​resource features to obtain sample predicted matching data, the sample predicted matching data including sample user matching data corresponding to each sample user information in the sample user set and / or sample resource matching data corresponding to each sample resource information in the sample resource set, wherein the sample user matching data is used to characterize the matching probability between the corresponding sample user information and each sample resource information in the sample resource set, and the sample resource matching data is used to characterize the matching probability between the corresponding sample resource information and each sample user information in the sample user set.

[0156] In one possible implementation, the model training module 912 includes:

[0157] The sample feature acquisition submodule 9121 is configured to input the sample user information into the user feature extraction network of the deep learning model, use the user feature extraction network to extract features of the sample user information, and obtain sample user features of the sample user information; input the sample resource information into the resource feature extraction network of the deep learning model, use the resource feature extraction network to extract features of the sample resource information, and obtain sample resource features of the sample resource information.

[0158] The splicing feature acquisition submodule 9122 is configured to input the sample user features of all sample user information in the sample user set into the user feature splicing network of the deep learning model, and use the user feature splicing network to splice the sample user features of all sample user information in the sample user set to obtain sample splicing user features; input the sample resource features of all sample resource information in the sample resource set into the resource feature splicing network of the deep learning model, and use the resource feature splicing network to splice the sample resource features of all sample resource information in the sample resource set to obtain sample splicing resource features.

[0159] The sample prediction matching data acquisition submodule 9123 is configured to input the sample splicing user features and the sample splicing resource features into the prediction network of the deep learning model, use the prediction network to match the sample splicing user features and the sample splicing resource features, and obtain the sample prediction matching data.

[0160] The sample identification determination submodule 9124 is configured to determine a sample user identification and a sample resource identification associated with each matching probability of each sample user matching data in the sample matching data.

[0161] The tag value acquisition submodule 9125 is configured to acquire a tag value corresponding to each matching probability according to the sample user identifier and the sample resource identifier associated with the matching probability.

[0162] The label data generation submodule 9126 is configured to generate label data based on the label value corresponding to each matching probability and the order of each matching probability in the sample matching data, wherein the order of each label value in the label data is the same as the order of the matching probability corresponding to each label value in the sample matching data.

[0163] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0164] Fig.10 1 is a block diagram of a server of a resource recall method according to an exemplary embodiment. The server 10 can be used to execute the steps executed by the server in the resource recall method. Fig.10The server 10 may include one or more (102a, 102b, ..., 102n are used to illustrate) processors 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, a power supply and / or a camera. It can be understood by those skilled in the art that Fig.10 The structure shown is only for illustration and does not limit the structure of the above electronic device. Fig.10 More or fewer components as shown, or with Fig.10 Different configurations are shown.

[0165] It should be noted that the one or more processors 102 and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the server 10 (or mobile device). As described in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0166] The memory 104 can be used to store software programs and modules of application software, such as program instructions / data storage devices corresponding to the method in the embodiment of the present invention. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, to implement the above-mentioned resource recall method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the server 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0167] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the server 10. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (Radio Frequency, RF) module, which is used to communicate with the Internet wirelessly.

[0168] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the server 10 (or mobile device).

[0169] In an exemplary embodiment, a non-transitory computer-readable storage medium is also provided. When instructions in the storage medium are executed by a processor of a computer device, the computer device is enabled to execute the steps executed by the server in the above-mentioned resource recall method.

[0170] In an exemplary embodiment, a computer program product is further provided. When instructions in the computer program product are executed by a processor of a computer device, the computer device is enabled to execute the steps executed by the server in the above resource recall method.

[0171] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0172] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A resource recall method, characterized in that: include: Acquire a user set and a resource set, wherein the user set includes user information of multiple users, and the resource set includes resource information of multiple resources to be delivered; splicing user features of all user information in the user set to obtain spliced ​​user features, and splicing resource features of all resource information in the resource set to obtain spliced ​​resource features; Performing matching processing on the spliced ​​user features and the spliced ​​resource features to obtain predicted matching data, wherein the predicted matching data includes user matching data corresponding to each user information in the user set and / or resource matching data corresponding to each resource information in the resource set, wherein the user matching data is used to represent the matching probability between the corresponding user information and each resource information in the resource set, and the resource matching data is used to represent the matching probability between the corresponding resource information and each user information in the user set; According to the predicted matching data, target resources that need to be recalled by users corresponding to each user information are screened out from the resource set.

2. The method according to claim 1, characterized in that The method further comprises: Feature extraction is performed on the user information and the resource information respectively to obtain user features of each user information in the user set and resource features of each resource information in the resource set.

3. The method according to claim 1, characterized in that The step of selecting target resources to be recalled by users corresponding to each user information from the resource set according to the predicted matching data includes: Determine a target matching probability from a plurality of matching probabilities included in each user matching data or a plurality of matching probabilities included in each resource matching data; The resource information and user information associated with the target matching probability are determined, and the resource information associated with the target matching probability is used as the target resource to be recalled by the user corresponding to the user information associated with the target matching probability.

4. The method according to claim 3, characterized in that The step of determining the target matching probability from the multiple matching probabilities included in each user matching data or the multiple matching probabilities included in each resource matching data includes: sorting the matching probabilities included in each user's matching data, and determining the target matching probability based on the sorting result; or, sorting the matching probabilities contained in each resource matching data, and determining the target matching probability based on the sorting result; or, Determine the target matching probability according to the comparison result between each matching probability included in each user matching data and a preset threshold; or, The target matching probability is determined based on the comparison result between each matching probability included in each resource matching data and a preset threshold.

5. The method according to claim 3, characterized in that: The determining of resource information and user information associated with the target matching probability includes: Obtaining a resource identifier and a user identifier corresponding to the target matching probability; determining resource information associated with the target matching probability according to the resource identifier, and determining user information associated with the target matching probability according to the user identifier, Among them, the user identifier is the input sequence information of the user feature input prediction model of the user information or the preset user number information, and the resource identifier is the input sequence information of the resource feature input prediction model of the resource information or the preset resource number information.

6. The method according to claim 2, characterized in that Inputting the user information in the user set and the resource information in the resource set into a prediction model, and obtaining the prediction matching data through the prediction model; The prediction model is trained by the following method: Acquire sample data, the sample data including a sample user set, a sample resource set and a label set, the sample user set including multiple sample user information, the sample resource set including multiple sample resource information, the label set including a label value corresponding to each sample user information or each sample resource information, each label value being used to indicate whether a piece of sample user information matches a piece of sample resource information; Inputting the sample user information and the sample resource information into a deep learning model, obtaining sample splicing user features and sample splicing resource features, performing matching processing on the sample splicing user features and the sample splicing resource features, and obtaining sample prediction matching data; Determining label data corresponding to the sample predicted matching data according to the sample predicted matching data and the label value in the sample data; Calculate the loss value based on the sample prediction matching data and the label data; The deep learning model is trained according to the loss value to obtain a prediction model.

7. The method according to claim 6, characterized in that The acquiring of sample splicing user features and sample splicing resource features, matching the sample splicing user features and the sample splicing resource features, and obtaining sample prediction matching data includes: splicing sample user features of all sample user information in the sample user set to obtain sample splicing user features, and splicing sample resource features of all sample resource information in the sample resource set to obtain sample splicing resource features; The sample splicing user features and the sample splicing resource features are matched to obtain sample prediction matching data, wherein the sample prediction matching data includes sample user matching data corresponding to each sample user information in the sample user set and / or sample resource matching data corresponding to each sample resource information in the sample resource set, wherein the sample user matching data is used to characterize the matching probability between the corresponding sample user information and each sample resource information in the sample resource set, and the sample resource matching data is used to characterize the matching probability between the corresponding sample resource information and each sample user information in the sample user set.

8. The method according to claim 7, characterized in that The method further comprises: Feature extraction is performed on the sample user information to obtain sample user features of the sample user information; feature extraction is performed on the sample resource information to obtain sample resource features of the sample resource information.

9. The method according to claim 6, characterized in that The determining, according to the sample predicted matching data and the label value in the sample data, label data corresponding to the sample predicted matching data comprises: Determining a sample user identifier and a sample resource identifier associated with each matching probability of each sample user matching data in the sample predicted matching data; Acquire a label value corresponding to each matching probability according to the sample user identifier and the sample resource identifier associated with the matching probability; Generate label data based on the label value corresponding to each matching probability and the ranking of each matching probability in the sample predicted matching data, wherein the ranking of each label value in the label data is the same as the ranking of the matching probability corresponding to each label value in the sample predicted matching data.

10. A resource recall device, characterized in that: include: An information acquisition unit is configured to acquire a user set and a resource set, wherein the user set includes user information of a plurality of users, and the resource set includes resource information of a plurality of resources to be delivered; The predicted matching data acquisition unit is configured to splice user features of all user information in the user set to obtain spliced ​​user features, and splice resource features of all resource information in the resource set to obtain spliced ​​resource features; match the spliced ​​user features and the spliced ​​resource features to obtain predicted matching data, wherein the predicted matching data includes user matching data corresponding to each user information in the user set and / or resource matching data corresponding to each resource information in the resource set, wherein the user matching data is used to represent the matching probability between the corresponding user information and each resource information in the resource set, and the resource matching data is used to represent the matching probability between the corresponding resource information and each user information in the user set; The target resource determination unit is configured to filter out target resources that need to be recalled by users corresponding to each user information from the resource set according to the predicted matching data.

11. The device according to claim 10, characterized in that The predicted matching data acquisition unit is configured to perform feature extraction on the user information and the resource information respectively, and obtain the user feature of each user information in the user set and the resource feature of each resource information in the resource set.

12. The device according to claim 10, characterized in that The target resource determination unit comprises: A target matching probability selection module is configured to determine a target matching probability from a plurality of matching probabilities included in each user matching data or a plurality of matching probabilities included in each resource matching data; The target resource determination module is configured to determine resource information and user information associated with the target matching probability, and use the resource information associated with the target matching probability as the target resource to be recalled by the user corresponding to the user information associated with the target matching probability.

13. The device according to claim 12, characterized in that The target matching probability selection module is configured as follows: sorting the matching probabilities included in each user's matching data, and determining the target matching probability based on the sorting result; or, sorting the matching probabilities contained in each resource matching data, and determining the target matching probability based on the sorting result; or, Determine the target matching probability according to the comparison result between each matching probability included in each user matching data and a preset threshold; or, The target matching probability is determined based on the comparison result between each matching probability included in each resource matching data and a preset threshold.

14. The device according to claim 11, characterized in that The target resource determination module includes: An identification acquisition submodule, configured to acquire a resource identification and a user identification corresponding to the target matching probability; A determination submodule is configured to determine the resource information associated with the target matching probability based on the resource identifier, and to determine the user information associated with the target matching probability based on the user identifier, wherein the user identifier is the input sequence information of the user feature input prediction model or the preset user number information of the user information, and the resource identifier is the input sequence information of the resource feature input prediction model or the preset resource number information.

15. The device according to claim 10, characterized in that The device further includes a model acquisition unit, which is used to process the user information in the user set and the resource information in the resource set to obtain the predicted matching data; The model acquisition unit comprises: a sample acquisition module configured to acquire sample data, wherein the sample data includes a sample user set, a sample resource set, and a label set, wherein the sample user set includes a plurality of sample user information, the sample resource set includes a plurality of sample resource information, and the label set includes a label value corresponding to each sample user information or each sample resource information, wherein each label value is used to indicate whether a piece of sample user information matches a piece of sample resource information; The model training module is configured to input the sample user information and the sample resource information into a deep learning model, obtain sample splicing user features and sample splicing resource features, match the sample splicing user features and the sample splicing resource features, and obtain sample prediction matching data; determine the label data corresponding to the sample prediction matching data according to the sample prediction matching data and the label value in the sample data; calculate the loss value according to the sample prediction matching data and the label data; and train the deep learning model according to the loss value to obtain a prediction model.

16. The device according to claim 15, characterized in that The model training module is configured as follows: splicing sample user features of all sample user information in the sample user set to obtain sample splicing user features, and splicing sample resource features of all sample resource information in the sample resource set to obtain sample splicing resource features; The sample splicing user features and the sample splicing resource features are matched to obtain sample prediction matching data, wherein the sample prediction matching data includes sample user matching data corresponding to each sample user information in the sample user set and / or sample resource matching data corresponding to each sample resource information in the sample resource set, wherein the sample user matching data is used to characterize the matching probability between the corresponding sample user information and each sample resource information in the sample resource set, and the sample resource matching data is used to characterize the matching probability between the corresponding sample resource information and each sample user information in the sample user set.

17. The device according to claim 16, characterized in that The model training module is configured to extract features from the sample user information to obtain sample user features of the sample user information; Feature extraction is performed on the sample resource information to obtain sample resource features of the sample resource information.

18. The device according to claim 15, characterized in that The model training module also includes: A sample identification determination submodule, configured to determine a sample user identification and a sample resource identification associated with each matching probability of each sample user matching data in the sample prediction matching data; A label value acquisition submodule is configured to acquire a label value corresponding to each matching probability according to a sample user identifier and a sample resource identifier associated with the matching probability; The label data generation submodule is configured to generate label data according to the label value corresponding to each matching probability and the ranking of each matching probability in the sample predicted matching data, wherein the ranking of each label value in the label data is the same as the ranking of the matching probability corresponding to each label value in the sample predicted matching data.

19. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the resource recall method as described in any one of claims 1 to 9.

20. A storage medium, characterized in that: When the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the resource recall method as described in any one of claims 1 to 9.

Citation Information

Patent Citations

  • A resource recommendation method and device

    CN109241431A

  • Information recommendation method and device

    CN110399550A

  • Resource recommendation method and device and computer storage medium

    CN110704739A

  • Method and system for recommending objects

    CN111159578A

  • Object recommendation method and device, computing equipment and storage medium

    CN111339240A