Data processing method, electronic device and storage medium

Abstract

The embodiment of the present application discloses a data processing method, an electronic device and a storage medium, which are applied to the field of machine learning technology. The method includes: obtaining a recommendation information set, generating a recommendation feature set according to the recommendation information set, determining the influencing factors of the object behavior feature and the influencing factors of the resource behavior feature according to the object attribute feature and the resource attribute feature, determining the target object behavior feature according to the influencing factor and the object behavior feature, and determining the target resource behavior feature according to the influencing factor and the resource behavior feature, determining the target probability of the target object responding to the recommended behavior of the recommended resource according to the object attribute feature, the resource attribute feature, the target object behavior feature and the target resource behavior feature, and pushing the recommended resource to the object terminal according to the target probability. By adopting the embodiment of the present application, the recommendation prediction efficiency and accuracy of the recommended resource can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of machine learning, and in particular to a data processing method, an electronic device, and a storage medium. Background Art

[0002] In the recommendation task, the main purpose is to predict the recommended resources that the user is interested in and push them to the user terminal to improve the effective recommendation rate of the recommendation task. For example, predict the short videos that the user is interested in and recommend them, so as to increase the click-through rate and viewing rate of the short videos. The existing prediction method usually builds a prediction model to jointly predict the user-related attributes and the attributes of the recommended resources. However, in this method, only the above two attributes are used to directly predict, so that there are fewer features involved in the prediction process, resulting in low recommendation prediction efficiency and accuracy. Therefore, how to improve the recommendation prediction efficiency and accuracy of recommended resources has become an urgent problem to be solved. Summary of the invention

[0003] The embodiments of the present application provide a data processing method, an electronic device, and a storage medium, which can effectively improve the recommendation prediction efficiency and accuracy of recommended resources.

[0004] On the one hand, an embodiment of the present application provides a data processing method, the method comprising:

[0005] Acquire a recommendation information set; the recommendation information set includes object attribute information and object behavior information of the target object, resource attribute information and resource behavior information of the resource to be recommended;

[0006] Generate a recommended feature set according to the recommended information set; the recommended feature set includes object attribute features of object attribute information, object behavior features of object behavior information, resource attribute features of resource attribute information, and resource behavior features of resource behavior information;

[0007] Determine the influencing factors of the object behavior characteristics and the influencing factors of the resource behavior characteristics according to the object attribute characteristics and the resource attribute characteristics;

[0008] Determine the target object behavior characteristics according to the influencing factors of the object behavior characteristics and the object behavior characteristics, and determine the target resource behavior characteristics according to the influencing factors of the resource behavior characteristics and the resource behavior characteristics;

[0009] The target probability of the target object responding to the recommended resource is determined according to the object attribute characteristics, resource attribute characteristics, target object behavior characteristics and target resource behavior characteristics, and the recommended resource is pushed to the target object's object terminal according to the target probability.

[0010] On the one hand, an embodiment of the present application provides a data processing device, the device comprising:

[0011] An acquisition module is used to acquire a recommendation information set; the recommendation information set includes object attribute information and object behavior information of the target object, resource attribute information and resource behavior information of the resource to be recommended;

[0012] A processing module, configured to generate a recommended feature set according to the recommended information set; the recommended feature set includes object attribute features of the object attribute information, object behavior features of the object behavior information, resource attribute features of the resource attribute information, and resource behavior features of the resource behavior information;

[0013] A determination module, used to determine the influencing factors of the object behavior characteristics and the influencing factors of the resource behavior characteristics according to the object attribute characteristics and the resource attribute characteristics;

[0014] The determination module is further used to determine the target object behavior characteristics according to the influencing factors of the object behavior characteristics and the object behavior characteristics, and to determine the target resource behavior characteristics according to the influencing factors of the resource behavior characteristics and the resource behavior characteristics;

[0015] The determination module is also used to determine the target probability of the target object responding to the recommended resource based on the object attribute characteristics, resource attribute characteristics, target object behavior characteristics and target resource behavior characteristics, and push the recommended resource to the object terminal of the target object according to the target probability.

[0016] On the one hand, an embodiment of the present application provides an electronic device, which includes a processor and a memory, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute some or all of the steps in the above method.

[0017] On the one hand, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program, and the computer program includes program instructions, which, when executed by a processor, are used to execute some or all of the steps in the above method.

[0018] Accordingly, according to one aspect of the present application, a computer program product or a computer program is provided, wherein the computer program product or the computer program comprises program instructions, and when the program instructions are executed by a processor, some or all of the steps in the above method are implemented.

[0019] In an embodiment of the present application, a recommendation feature set can be generated based on a recommendation information set, an influencing factor of an object behavior feature and an influencing factor of a resource behavior feature can be determined based on an object attribute feature and a resource attribute feature, a target object behavior feature can be determined based on an influencing factor of the object behavior feature and the object behavior feature, a target resource behavior feature can be determined based on an influencing factor of the resource behavior feature and the resource behavior feature, a target probability of a target object responding to a recommended behavior of a recommended resource can be determined based on the object attribute feature, the resource attribute feature, the target object behavior feature and the target resource behavior feature, and the recommended resource can be pushed to the object terminal according to the target probability. Through the above method, not only can the target probability be determined in combination with more feature information, but also feature selection can be performed on the object behavior feature and the resource behavior feature through the influencing factor to determine a target behavior feature that is more suitable in the prediction process, thereby improving the recommendation prediction efficiency and accuracy of the recommended resource, and thus facilitating improving the effect of subsequent recommendations based on the target probability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1 A schematic diagram of an application architecture provided for an embodiment of the present application;

[0022] Figure 2 A flowchart of a data processing method provided in an embodiment of the present application;

[0023] Figure 3a A schematic diagram of a resource to be recommended provided in an embodiment of the present application;

[0024] Figure 3b A schematic diagram of a resource to be recommended provided in an embodiment of the present application;

[0025] Figure 3c A schematic diagram of a resource to be recommended provided in an embodiment of the present application;

[0026] Figure 4 A flowchart of a data processing method provided in an embodiment of the present application;

[0027] Figure 5a A schematic diagram of a scenario for determining a recommended feature set provided in an embodiment of the present application;

[0028] Figure 5b A schematic diagram of a scenario for determining a recommended feature set provided in an embodiment of the present application;

[0029] Figure 5c A schematic diagram of a scenario for determining a recommended feature set provided in an embodiment of the present application;

[0030] Figure 6a An example diagram of a feature extraction scenario based on a target prediction model provided in an embodiment of the present application;

[0031] Figure 6b An example diagram of a feature extraction scenario based on a target prediction model provided in an embodiment of the present application;

[0032] Figure 6c An example diagram of a feature extraction scenario based on a target prediction model provided in an embodiment of the present application;

[0033] Figure 7 A schematic diagram of a push scenario of resources to be recommended based on a target prediction model provided in an embodiment of the present application;

[0034] Figure 8 A flowchart of a data processing method provided in an embodiment of the present application;

[0035] Figure 9a A schematic diagram of a model training scenario provided in an embodiment of the present application;

[0036] Figure 9b A schematic diagram of a prediction model provided in an embodiment of the present application;

[0037] Fig.10 A schematic diagram of the structure of a data processing device provided in an embodiment of the present application;

[0038] Fig.11 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0040] The data processing method proposed in the embodiment of the present application is implemented in an electronic device, which can be a server or a terminal. Among them, the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. The terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, an intelligent voice interaction device, a smart home appliance, a vehicle terminal, an aircraft, etc., but is not limited to this.

[0041] In some embodiments, see Figure 1 , Figure 1 The following is a schematic diagram of an application architecture provided in an embodiment of the present application. The data processing method proposed in the present application can be executed through the application architecture. Figure 1 As shown, it can include an electronic device and an object terminal of a target object; wherein the target object can be a user (such as a user) of a target application; the electronic device can obtain a recommendation information set and generate a recommendation feature set according to the target object and the resource to be recommended, the recommendation feature set includes relevant recommendation features of the target object (such as object attribute features and object behavior features) and relevant recommendation features of the resource to be recommended (such as resource attribute features and resource behavior features), and feature selection is performed according to the recommendation feature set to determine the target recommendation feature set, such as feature selection of object behavior features and resource behavior features according to object attribute features and resource attribute features to determine target object behavior features and target resource behavior features, the target recommendation feature set includes object attribute features, resource attribute features, target object behavior features and target resource behavior features, and the target probability of the target object responding to the recommended behavior to the recommended resource is determined according to the target recommendation feature set; subsequently, the recommendation strategy for the recommended resource can be determined according to the target probability, and the recommendation strategy can be pushed based on the recommendation strategy. Optionally, the above process can also be performed through a target prediction model, such as a target prediction model can be deployed in the electronic device, in which a recommendation feature set is generated through a recommendation information set, and a target probability is obtained according to the recommendation feature set.

[0042] Understandably, Figure 1 It is only an exemplary representation of the possible application architecture of the technical solution of the present application, and does not limit the specific architecture of the technical solution of the present application, that is, the technical solution of the present application can also provide other forms of application architecture.

[0043] Optionally, in some embodiments, the electronic device may execute the data processing method to improve the prediction efficiency and accuracy of the recommended resources according to actual business needs. That is, the electronic device may obtain the corresponding recommendation information set according to the target object and the resource to be recommended to generate a recommendation feature set, perform feature selection on the behavior features (object behavior features and resource behavior features) according to the attribute features (object attribute features and resource attribute features) in the recommendation feature set, obtain the target behavior features that are more suitable in the prediction process, and determine the target probability according to the attribute features and the target behavior features, and recommend the resource to be recommended according to the target probability, so as to achieve accurate push and improve the recommendation effect.

[0044] Among them, the resource to be recommended can be a product that can be recommended in the target application. A target application can have multiple types of resources to be recommended. The different types of resources to be recommended can belong to different domains, and the different types of resources to be recommended have different product scenarios. Therefore, the technical solution of the present application can realize the joint recommendation of multiple resources to be recommended (also known as cross-domain recommendation). For example, the resource to be recommended can be a short video that can be recommended in the target application, and the product scenario can be the display page of the short video recommendation channel; for another example, the resource to be recommended can be a live broadcast room that can be recommended in the target application, and the product scenario can be the application homepage of the target application, etc.

[0045] Therefore, the technical solution of the present application can be applied to any recommendation task. For example, it can be applied to the recommendation task of social applications. At this time, the resources to be recommended can be the live broadcast room of the online anchor, the short videos that have been released, etc., and the target objects can be users who use social applications. The electronic device can determine the user's target probability for the product (live broadcast room or short video, etc.) based on the technical solution of the present application, and push the product to the user's user terminal based on the target probability to increase the user's click-through rate or viewing time of the recommended product. For another example, it can also be applied to the recommendation task of shopping applications. At this time, the resources to be recommended can be commodities, commodity usage notes posted by online users, etc., and the target objects can be users who use shopping applications. The electronic device can determine the user's target probability for the product (commodity or commodity usage notes) based on the technical solution of the present application, and push the product to the user's user terminal based on the target probability to increase the user's click-through rate and purchase rate of the recommended product.

[0046] Optionally, the data involved in this application, such as a recommendation information set, etc., can be stored in a database, or can be stored in a blockchain, such as stored in a blockchain distributed system, which is not limited in this application.

[0047] It is understood that the above scenarios are only examples and do not constitute a limitation on the application scenarios of the technical solutions provided in the embodiments of the present application. The technical solutions of the present application can also be applied to other scenarios. For example, it is known to those skilled in the art that with the evolution of the system architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0048] Based on the above description, the present application embodiment proposes a data processing method, which can be executed by the electronic device mentioned above. Figure 2 , Figure 2 A flow chart of a data processing method provided in an embodiment of the present application. Figure 2 As shown, the process of the data processing method of the embodiment of the present application may include the following:

[0049] S201: Obtain a set of recommended information.

[0050] The recommendation information set may include object attribute information and object behavior information of the target object, and resource attribute information and resource behavior information of the resource to be recommended.

[0051] In some embodiments, the target object may be a user who needs to be predicted for recommendation, and the object attribute information may include one or more user attributes, which can be used to characterize the relevant user information of the user. For example, the object attribute information may be the basic portrait attribute of the target object, such as the user's age, gender, and education level. The object behavior characteristics may include one or more behavior information generated for the resource to be recommended under the target object dimension. For example, the resource to be recommended is a short video product, so the object behavior characteristics may include the user's statistical behavior for watching short videos, such as the total duration of the user's short video watching within one day or three days, the user's short video like sequence within three days or one week, the user's short video collection sequence within three days or one week, and so on; wherein, any of the above sequences can be determined based on the short video identification, that is, the identification of each published short video can be generated according to the generation rules and according to the type, release time, etc., thereby obtaining the identification of the short video associated with the user (such as the short video liked by the user or the short video collected) to generate the relevant sequence. The generation rules can be set by relevant business personnel according to experience values. If the recommendation information set corresponding to different resources to be recommended contains sequences, there may be different generation methods, which can be set by relevant business personnel and are not limited here.

[0052] In some embodiments, the resource to be recommended may be a recommended product in the target application, such as a short video, a live broadcast room of an anchor, etc. Resource attribute information may include one or more resource attributes, which may be used to characterize the relevant object information of the resource or the object to which the resource belongs. For example, if the resource to be recommended is a live broadcast room (or video), and the object to which the resource belongs is the corresponding anchor (or the corresponding publisher; in different product scenarios, the same anchor may also be the publisher), the resource attribute information may be the basic portrait attributes of the anchor (or publisher), such as the age, gender, and education level of the anchor (or publisher); for another example, if the resource to be recommended is a commodity (or commodity usage notes), the resource attribute information may be the basic portrait attributes of the commodity (or the commodity indicated by the commodity usage notes; in different product scenarios, the same commodity may also be the commodity indicated in the commodity usage notes), such as the type of commodity, the praise rate of the store to which the commodity belongs, etc. Resource behavior characteristics may include one or more behavior information generated for the resource to be recommended under the dimension of the resource to be recommended. For example, the resource to be recommended is a short video product, so the resource behavior characteristics may include the statistical behavior of the short video for the user, such as the total number of times the short video is liked by the user within one day or three days, the user like sequence of the short video within three days or one week, the user collection sequence of the short video within three days or one week, etc. The specific information type contained in the above recommendation information set can be set by relevant business personnel according to the recommendation task of actual application, and is not limited here. Among them, the relevant description of the generated sequence can refer to the above description.

[0053] It should be noted that in the specific implementation of this application, when it involves user information such as user age and other related data, or any related data in any scenario where data needs to be obtained, when the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0054] In some embodiments, the resource to be recommended belongs to the target application in the target terminal, and the resource to be recommended can be any type of product that can be recommended in the target application. The product scenarios of different resources to be recommended can be the same or different, and the product scenario can refer to the application interface in the target application. In an application interface, only one resource to be recommended can be recommended, or multiple resources to be recommended can be recommended at the same time. In addition, there can be one or more product scenarios for the same resource to be recommended, that is, there can be multiple application interfaces in the target application that can recommend the same resource to be recommended; for the same resource to be recommended, if it is in different product scenarios, the recommendation prediction method proposed in this application can be used.

[0055] Therefore, the present application can jointly recommend different products in the target application, and the prediction resources used in the joint recommendation (such as relevant parameters involved in the process, etc.) can be partially shared and partially unique, so that the technical solution of the present application can combine more feature information when making recommendation predictions. Compared with modeling each resource to be recommended separately for recommendation prediction, this can combine the common features and difference features of the recommendation predictions of multiple resources to be recommended to improve the effect of recommendation prediction. The electronic device can be a background device to which the target application belongs. When the object terminal detects that the target application displays an application interface that can recommend products, it generates a recommendation instruction and sends it to the electronic device. When the electronic device receives the recommendation instruction, it can determine the resource type of the resource to be recommended based on the product scenario represented by the application interface, and obtain the recommendation information set corresponding to the target object and the resource to be recommended.

[0056] Among them, the resource to be recommended obtained by the electronic device may be one or more. When there are multiple resources to be recommended, the electronic device may construct a recommendation information set for each resource to be recommended and the target object, respectively, and each recommendation information set includes the object attribute information and object behavior information of the target object, and the resource attribute information and resource behavior information of a resource to be recommended. And when there are multiple resources to be recommended, the resource types of the multiple resources to be recommended may be different, and the different types of resources to be recommended belong to different recommendation tasks, and the results of recommendation prediction may be independent of each other. The process and principle of recommendation prediction for each recommendation information set are the same. When describing the prediction process below, one recommendation information set is taken as an example.

[0057] For example, Figure 3a-3c As shown, Figure 3a-3c A schematic diagram of a resource to be recommended provided in an embodiment of the present application; wherein the resource to be recommended may be a live broadcast room in a target application. In this case, the product scenario of the resource to be recommended in the target application may be the application homepage of the target application. The application homepage includes a live broadcast room recommendation column. In this case, the live broadcast room is represented as a live broadcast tag and is displayed in thumbnail form on the live broadcast room recommendation column, such as Figure 3a As shown; the product scenario can also be the live broadcast recommendation page of the target application, which directly displays the live broadcast room details screen, and the target object can enter the recommended different live broadcast rooms by sliding up and down, such as Figure 3b As shown, one slide indicates a recommendation of a live broadcast room;

[0058] For another example, the resource to be recommended may be a short video in the target application. In this case, the product scenario of the resource to be recommended in the target application may be the application homepage of the target application. The application homepage includes a short video recommendation column. In this case, the short video is represented by a short video tag and is displayed in a thumbnail in the short video recommendation column, such as Figure 3aAs shown; the product scenario can also be a short video recommendation page of the target application, which directly displays the short video details screen, and the target object can enter different recommended short videos by sliding up and down, such as Figure 3c As shown, one slide indicates a short video recommendation;

[0059] Among them, the above Figure 3a The indicated product scenario can recommend two different resources to be recommended, namely live broadcast rooms and short videos, at the same time, which means that the resources to be recommended have the same product scenario; at this time, the resources to be recommended obtained are products of different resource types (such as multiple live broadcast rooms and multiple short videos), and correspond to the recommendation tasks for live broadcast rooms and the recommendation tasks for short videos. The two recommendation tasks can be carried out simultaneously, and the recommendation results are independent of each other.

[0060] It is understandable that the technical solution of the present application can be applied to recommendation tasks of any application. Depending on the recommendation tasks, the target objects and the resources to be recommended may be different, and no limitation is made here.

[0061] S202: Generate a recommended feature set according to the recommended information set.

[0062] The recommended feature set includes object attribute features of object attribute information, object behavior features of object behavior information, resource attribute features of resource attribute information, and resource behavior features of resource behavior information.

[0063] In a possible implementation, the generation process and principle of each recommended feature in the recommended feature set are the same. Specifically, the recommended information is feature encoded according to the matching encoding rule to obtain an encoded feature vector, and the encoded feature vector is used as the corresponding recommended feature. Different recommended information may correspond to different encoding rules, which can be set by relevant business personnel based on experience.

[0064] In some embodiments, taking the age attribute in the object attribute information as an example, generating the corresponding object attribute feature can specifically be to divide the age attribute into multiple attribute categories, and encode according to the multiple attribute categories according to the encoding rules. For example, the age attribute is divided into multiple attribute categories [<18, 19-30, 31-40, 41-50, 51-60, >60]. If the age of the target object is 24, it means that the target attribute category divided is [19-30]. Therefore, the encoding feature vector obtained by feature encoding according to the encoding rules can be expressed as [0, 1, 0, 0, 0, 0]. Among them, the recommended information in the recommended information set except for the sequence type can generate corresponding recommendation features according to the above-described process and principle.

[0065] In some embodiments, the recommendation information may include recommendation information of a sequence type. Taking the short video like sequence in the object behavior information as an example, generating the corresponding object behavior feature may specifically be to divide the short video into multiple attribute categories and encode according to the multiple attribute categories according to the encoding rules. For example, the short video is divided into multiple attribute categories according to the publisher of the short video: the short video published by user A is encoded as 1, the short video published by user B is encoded as 2....; if the short video like sequence is: [short video A identifier, short video B identifier, short video C identifier, ....], and suppose that short video A is published by user A, short video B is published by user B, and short video C is published by user B, then the encoding feature vector obtained by feature encoding according to the encoding rules can be expressed as [1,2,2,...]. Furthermore, when the sequence is of numerical type, the sequence can also be directly used as the corresponding recommendation feature. Among them, the recommendation information of the sequence type in the recommendation information set can generate corresponding recommendation features according to the above-described process and principle.

[0066] In different recommendation tasks, the attribute classification method for the same recommendation information can be different. For example, in the recommendation task for the live broadcast room, the age attribute can be divided into [<18, 19-30, 31-40, 41-50, 51-60, >60], and in the recommendation task for short videos, the age attribute can be divided into [<20, 21-40, 41-45, 46-50, 51-60, >60], etc. The classification method can be set by relevant business personnel.

[0067] Therefore, corresponding coding feature vectors can be generated respectively from one or more attributes contained in the object attribute information / one or more behavior information contained in the object behavior information / one or more attributes contained in the resource attribute information / one or more behavior information contained in the resource behavior information, and the object attribute feature / object behavior feature / resource attribute feature / resource behavior feature can be composed of the one or more coding feature vectors. The specific method for an electronic device to generate resource attribute features and resource behavior features can be the same as the specific method for generating object attribute features and object behavior features, which will not be repeated here.

[0068] In some embodiments, when the electronic device obtains the encoded feature vector corresponding to each recommendation information in the recommendation information set, it can also perform feature embedding processing on the encoded feature vector to obtain an embedded feature vector corresponding to each recommendation information, and use the embedded feature vector as the corresponding recommendation feature to obtain a recommendation feature set. The above specific execution process can refer to the relevant description of the following embodiment.

[0069] S203: Determine the influencing factors of the object behavior characteristics and the influencing factors of the resource behavior characteristics according to the object attribute characteristics and the resource attribute characteristics.

[0070] In one possible implementation, the impact factor can be obtained by a target prediction model. The target prediction model can realize the joint prediction of cross-domain recommendation tasks. The target prediction model can include an impact factor prediction network, in which the impact factor of each behavior feature is generated according to the splicing features of the above two attribute features. Among them, when the impact factor prediction is performed on each recommended feature set, the model parameters of the impact factor prediction network are prediction resources shared by the common features in each recommended feature set, so that the information learned in different domains can be migrated to the impact factor prediction network associated with the common features, thereby realizing the use of the impact factor prediction network to generate the impact factor of each difference feature, and also realizing that the information learned in domain A can act on the recommendation prediction under domain B, so that when the recommendation prediction is performed on the user in the recommendation task indicated by domain B, if there is less information that can participate in the prediction (such as the behavior information generated by the user on the recommended resources indicated by domain B), it can be combined with the information transmitted in the recommendation task indicated by domain A for joint prediction. The target prediction model is obtained by training based on the process described in steps S801-S804 in the following embodiment.

[0071] In some embodiments, the impact factor prediction network can be composed of a fully connected layer and a gated network, and the number of fully connected layers can be set by relevant business personnel based on experience. The impact factor can be obtained by the gated network using the idea of ​​activation function through the following formula:

[0072]

[0073] Among them, x i is the i-th element value of the vector output by the object attribute features and resource attribute features through the fully connected layer, f(x i ) represents the influencing factor of the i-th behavior feature in the behavior features (object behavior features and resource behavior features).

[0074] In one possible implementation, since the above-mentioned object attribute features are features that characterize the user who needs to be recommended, the user can make recommendation predictions for different resources to be recommended in different product scenarios. For example, when the electronic device recommends a live broadcast room to the user, it will obtain the user's specified object attribute information to generate object attribute features for the live broadcast room, a resource to be recommended. When recommending a short video to the user, it will also obtain this part of the specified object attribute information to generate object attribute features for the short video, a resource to be recommended. Therefore, the object attribute feature can be used as a common feature for joint recommendation of multiple resources to be recommended. And the resource attribute feature is a feature that characterizes the resource to be recommended or the object to which the resource to be recommended belongs. Different resources to be recommended or the objects to which the resource to be recommended belongs can overlap in different product scenarios.

[0075] For example, when recommending a live broadcast room, the specified attribute information of anchor A corresponding to live broadcast room A is obtained to generate resource attribute features for the live broadcast room, a resource to be recommended; when recommending a short video, the specified attribute information of publisher A corresponding to short video A is obtained to generate resource attribute features for the short video, a resource to be recommended; if anchor A and publisher A are the same object, the obtained specified attribute information is the same, so the resource attribute feature can be used as a common feature for joint recommendation of multiple resources to be recommended, and the remaining recommendation features can be used as differential features for joint recommendation of multiple resources to be recommended. Therefore, the technical solution of the present application can be applied to cross-domain recommendation tasks that combine multiple overlapping recommendation resources (i.e., there are common features, such as object attribute information and resource attribute information).

[0076] In some embodiments, the electronic device can perform feature selection on the difference features (object behavior features and resource behavior features) based on the common features (object attribute features and resource attribute features) to determine the feature information to be extracted from the difference features, and the extracted feature information is the more suitable behavior feature to participate in the prediction process of the recommended resource. Therefore, the electronic device can determine the influencing factor of the object behavior feature and the influencing factor of the resource behavior feature based on the aforementioned two attribute features to achieve feature selection of the difference features based on the common features. The influence factor can be used to measure the role played by the corresponding behavior feature in predicting the target probability under the aforementioned two attribute feature dimensions. The larger the influence factor, the greater the role played, and the more feature information is extracted from the corresponding behavior feature; conversely, the smaller the influence factor, the smaller the role played, and the less feature information is extracted from the corresponding behavior feature. Thereby, the correction of the features involved in the prediction can be achieved, so that more attention can be paid to important feature information during the prediction, thereby improving the prediction accuracy.

[0077] For example, the object behavior feature is the total viewing time on the short video, and the impact factor is 0.7; the resource behavior feature is the total number of likes for short video A, and the impact factor is 0.3. This means that when predicting the target probability, the result of the object behavior feature has a greater impact on the prediction result, so more feature information can be extracted from the object behavior feature; the result of the resource behavior feature has a smaller impact on the prediction result, so less feature information can be extracted from the resource behavior feature.

[0078] S204: Determine the target object behavior characteristics according to the influencing factors of the object behavior characteristics and the object behavior characteristics, and determine the target resource behavior characteristics according to the influencing factors of the resource behavior characteristics and the resource behavior characteristics.

[0079] In some embodiments, determining the target object behavior feature according to the influencing factor of the object behavior feature and the object behavior feature may be to use the product of the influencing factor and the object behavior feature as the target object behavior feature. The target object behavior feature represents the adaptation feature information extracted from the object behavior feature.

[0080] In some embodiments, determining the target resource behavior feature according to the impact factor of the resource behavior feature and the resource behavior feature may be to use the product of the impact factor and the resource behavior feature as the target resource behavior feature. The target resource behavior feature represents the adaptation feature information extracted from the resource behavior feature.

[0081] S205 , determining a target probability of the target object responding to the recommended resource according to the object attribute characteristics, resource attribute characteristics, target object behavior characteristics and target resource behavior characteristics, and pushing the recommended resource to the target object's object terminal according to the target probability.

[0082] In some embodiments, the electronic device may determine the target probability through the above-mentioned target prediction model. The target prediction model may include a probability prediction network for the target probability. The target probability may be predicted by the probability prediction network based on object attribute characteristics, resource attribute characteristics, target object behavior characteristics, and target resource behavior characteristics. The target probability represents the probability that the target object responds to the recommended resource with respect to the recommended behavior. Among them, the specific type of the recommended behavior may vary depending on the different resources to be recommended and the different product scenarios in which the resources to be recommended are currently located. The recommended behavior may represent that the target object has generated an effective recommendation for the resource to be recommended.

[0083] For example, the resource to be recommended is a live broadcast room, and is in Figure 3a In the product scenario shown, the recommended behavior at this time can be a click behavior on the live broadcast room, and the target probability at this time can represent the predicted click rate of the target object for the recommended live broadcast room; for another example, the resource to be recommended is a live broadcast room, and it is in the following Figure 3bIn the product scenario shown, the recommended behavior at this time can be the behavior of watching the live broadcast room. If the playback screen of the currently recommended live broadcast room is watched for 3 seconds, it means an effective recommendation. The target probability at this time can represent the predicted viewing rate of the target object for the recommended live broadcast room.

[0084] In some embodiments, the target probability indicates the target object's interest in the recommended resource. Therefore, a recommendation strategy for the recommended resource can be determined based on the target probability, and the recommended resource can be pushed to the target terminal according to the recommendation strategy. The recommendation strategy can be set by relevant business personnel. For example, the recommendation strategy can be to push the recommended resources to the target terminal in order from large to small in the target probability; for another example, the recommendation strategy can be to push the recommended resources whose target probability is greater than or equal to a preset probability threshold in sequence; and so on.

[0085] In an embodiment of the present application, a recommendation information set can be obtained, a recommendation feature set can be generated based on the recommendation information set, an influencing factor of an object behavior feature and an influencing factor of a resource behavior feature can be determined based on an object attribute feature and a resource attribute feature, a target object behavior feature can be determined based on the influencing factor of the object behavior feature and the object behavior feature, and a target resource behavior feature can be determined based on the influencing factor of the resource behavior feature and the resource behavior feature, a target probability of the target object responding to the recommended behavior of the recommended resource can be determined based on the object attribute feature, the resource attribute feature, the target object behavior feature and the target resource behavior feature, and the recommended resource can be pushed to the object terminal of the target object according to the target probability. Through the above method, not only can the target probability be determined in combination with more feature information, but also feature selection can be performed on the object behavior feature and the resource behavior feature through the influencing factor to determine the target behavior feature that is more suitable in the prediction process, thereby improving the recommendation prediction efficiency and accuracy of the recommended resource, which is beneficial to improving the effect of subsequent recommendations based on the target probability.

[0086] See also Figure 4 , Figure 4 The following is a flow chart of a data processing method provided in an embodiment of the present application, which can be executed by the electronic device mentioned above. Figure 4 As shown, the process of the data processing method in the embodiment of the present application may include the following:

[0087] S401, obtaining a set of recommended information. The specific implementation of step S401 can refer to the relevant description of the above embodiment, which will not be repeated here.

[0088] S402: Generate a recommended feature set according to the recommended information set.

[0089] The recommended feature set includes object attribute features, object behavior features, resource attribute features, and resource behavior features.

[0090] In a possible implementation, the generation process and principle of each recommended feature in the recommended feature set are the same. Specifically, the recommended information is feature encoded according to the matching encoding rule to obtain the encoded feature vector, the embedded feature matrix corresponding to the recommended information is obtained, and the embedded feature vector corresponding to each recommended information is determined according to the encoded feature vector and the embedded feature matrix, and the embedded feature vector is used as the corresponding recommended feature. The specific method of feature encoding can refer to the relevant description of the above embodiment.

[0091] In some embodiments, the corresponding embedded feature vector is determined according to the encoded feature vector and the embedded feature matrix, specifically, the encoded feature vector is multiplied with each column of the embedded feature matrix in turn and summed to obtain the calculation result corresponding to each column, and the vector obtained by splicing the calculation results corresponding to each column is used as the embedded feature vector. For example, the encoded feature vector is [1,2,3], and the element values ​​of each column of the embedded feature matrix are [(1,2,3), (2,3,1), (3,4,5)] in turn, so the calculation results corresponding to each column are: 1*1+2*2+3*3=14, 1*2+2*3+3*1=9, 1*3+2*4+3*5=26; therefore, the spliced ​​vector is [14,9,26].

[0092] Furthermore, in a recommendation task, the embedding feature matrices corresponding to each recommendation information may be the same or different. Between different recommendation tasks, the embedding feature matrices corresponding to the same recommendation information may be completely or partially the same or completely different.

[0093] For example, Figure 5a-5c As shown, Figure 5a-5c A schematic diagram of a scenario for determining a recommended feature set provided by an embodiment of the present application; wherein, assuming that recommendation task 1 is a recommendation task for a live broadcast room, recommendation information set 1 corresponding to recommendation task 1 includes object attribute information (including attribute A and attribute B), object behavior information (including behavior A and behavior B), resource attribute information (including attribute C and attribute D), and resource behavior information (including behavior C and behavior D); assuming that recommendation task 2 is a recommendation task for a short video, recommendation information set 2 corresponding to recommendation task 2 includes object attribute information (including attribute A and attribute B), object behavior information (including behavior A and behavior B), resource attribute information (including attribute C and attribute D), and resource behavior information (including behavior C and behavior D);

[0094] like Figure 5a, the embedding feature matrix corresponding to each recommendation information in recommendation task 1 is the same, that is, recommendation information set 1 obtains the corresponding recommendation feature set 1 based on matrix 1; if the embedding feature matrix used between recommendation task 1 and recommendation task 2 is the same, then recommendation information set 2 also obtains the corresponding recommendation feature set based on matrix 1; or, if the embedding feature matrix used between recommendation task 1 and recommendation task 2 is different, then recommendation information set 2 can obtain the corresponding recommendation feature set 2 based on matrix 2;

[0095] like Figure 5b , the embedding feature matrix corresponding to each recommendation information in recommendation task 1 is different, that is, recommendation information set 1 obtains the corresponding recommendation feature set 1 based on matrices 1-8; if the embedding feature matrices used between recommendation task 1 and recommendation task 2 are all the same, then recommendation information set 2 can also obtain the corresponding recommendation feature set 2 based on matrices 1-8; or, if the embedding feature matrices used between recommendation task 1 and recommendation task 2 are all different, then recommendation information set 2 can obtain the corresponding recommendation feature set based on matrices 9-16;

[0096] like Figure 5c , the recommended information set 1 obtains the corresponding recommended feature set 1 based on matrix 1-8; if the embedded feature matrix used between the recommended task 1 and the recommended task 2 is partially the same, for example, the recommended information using the same embedded feature matrix is ​​the object attribute information and the resource attribute information (i.e., the common features between cross-domain recommendation tasks), then the attributes A, B, C and D in the recommended information set 2 use the same embedded feature matrix (matrix 1-4) as the recommended information set 1, and the behaviors A, B, C and D in the recommended information set 2 use matrices 9-12; among which, if the same embedded feature matrix is ​​used between multiple recommendation tasks, then when obtaining the embedded feature matrix corresponding to each recommended information, the recommended information needs to be aligned, that is, the attribute A in the recommended information set 1 is aligned with the attribute A in the recommended information set 2, the attribute B in the recommended information set 1 is aligned with the attribute B in the recommended information set 2, the attribute C in the recommended information set 1 is aligned with the attribute C in the recommended information set 2, and the attribute D in the recommended information set 1 is aligned with the attribute D in the recommended information set 2.

[0097] In some embodiments, the above-mentioned embedded feature matrix can be set by relevant business personnel based on experience values. If there is aligned recommendation information between different recommendation tasks, the same embedded feature matrix is ​​set for the aligned recommendation information. It can also be obtained by training as a model parameter of the target prediction model. If there is aligned recommendation information between different recommendation tasks, the aligned recommendation information is trained to obtain the same embedded feature matrix. That is, the above process can be obtained through the target prediction model. The target prediction model may include a feature generation layer, and the feature generation layer generates a recommendation feature set based on the recommendation information set. The embedded feature matrix can be used as a model parameter in the target prediction model. At this time, the input recommendation information set may include a resource type identifier corresponding to the resource to be recommended (such as Figure 5a The target prediction model can select the corresponding embedded feature matrix for the recommended information set based on the resource type identifier to generate a recommended feature set. The target prediction model is trained based on the process described in steps S801-S804 in the following embodiment.

[0098] S403: Determine the influencing factors of the object behavior characteristics and the influencing factors of the resource behavior characteristics according to the object attribute characteristics and the resource attribute characteristics.

[0099] In a possible implementation, the electronic device may perform feature selection on the object behavior features and the resource behavior features according to the object attribute features and the resource attribute features to obtain an impact factor, and extract appropriate feature information through the impact factor for subsequent target probability prediction.

[0100] In some embodiments, the target prediction model may include an impact factor prediction network. The impact factor prediction network may be used to predict the impact factor. Determining the impact factor of the object behavior feature and the impact factor of the resource behavior feature based on the object attribute feature and the resource attribute feature may specifically be to fuse the object attribute feature and the resource attribute feature to obtain a fused attribute feature, and the impact factor prediction network generates the impact factor of the object behavior feature and the impact factor of the resource behavior feature based on the fused attribute feature.

[0101] In some embodiments, the object attribute features and resource attribute features are fused to obtain the fused attribute features, which may be specifically, the object attribute features and resource attribute features are taken as a feature set, and feature crossover is performed on every two features in the feature set (such as calculating the inner product of every two vectors), to obtain the crossover feature values ​​corresponding to every two features, and the crossover feature values ​​corresponding to every two features are sequentially spliced ​​to obtain the fused attribute features. That is, feature crossover can be expressed as:

[0102] <V i ,V j >1≤i≤N,1≤j≤N

[0103] Among them, <> represents the vector inner product, V i ,V j Represents the i-th feature and the j-th feature in the feature set, i can be equal to j, and N represents the number of features in the feature set.

[0104] It can be understood that the fused attribute feature can include feature association information between object attribute features, between resource attribute features, and between object attribute features and resource attribute features, which represents the relationship between features obtained through feature interaction between multiple features, thereby enabling the fused attribute feature to more deeply combine the relationship between two features and enrich the feature representation.

[0105] There are other ways to fuse features, which are not limited here. The specific process of generating the influencing factors of object behavior features and resource behavior features by the influencing factor prediction network based on the fused attribute features can be the same as the specific process of generating the influencing factors by the influencing factor prediction network based on the object attribute features and resource attribute features in the above embodiment.

[0106] S404: extracting features from the object behavior features to obtain object extraction features of the object behavior features, and extracting features from the resource behavior features to obtain resource extraction features of the resource behavior features.

[0107] In a possible implementation, the number of object behavior features may be N, where N is a positive integer greater than 1. The electronic device extracts features from the object behavior features to obtain an object extraction feature of the object behavior features, which may be, based on the N object behavior features, extracting features from the i-th object behavior feature to obtain an object extraction feature of the i-th object behavior feature. The i-th object behavior feature is any one of the N object behavior features, where i is a positive integer less than or equal to N. The object extraction feature corresponding to each object behavior feature is obtained based on the corresponding object behavior feature and the remaining object behavior features, so that the object extraction feature can cover more feature information. Similarly, the number of resource behavior features may be Q, where Q is a positive integer greater than 1. The electronic device extracts features from the resource behavior features to obtain a resource extraction feature of the resource behavior features, which may be, based on the Q resource behavior features, extracting features from the q-th resource behavior feature to obtain a resource extraction feature of the q-th resource behavior feature. The q-th resource behavior feature is any one of the Q resource behavior features, where q is a positive integer less than or equal to Q. The resource extraction feature corresponding to each resource behavior feature is obtained based on the corresponding resource behavior feature and other resource behavior features, so that the resource extraction feature can cover more feature information.

[0108] In one possible implementation, the process and principle of feature extraction for each object behavior feature or each resource behavior feature are the same. Here, the feature extraction process of the i-th object behavior feature is used as an example for explanation. The electronic device can be implemented by a target prediction model, and the target prediction model can include a feature extraction layer, and the feature extraction layer can be used to extract features of N object behavior features to obtain object extraction features corresponding to each object behavior feature. Among them, feature extraction through the feature extraction layer can specifically be to determine the transition behavior feature of each object behavior feature in the feature extraction layer, determine the degree of correlation between the i-th object behavior feature and each object behavior feature based on the transition behavior feature, and determine the feature information extracted from each object behavior feature under the i-th object behavior feature according to the correlation and the transition behavior feature, and determine the object extraction feature of the i-th object behavior feature according to the extracted feature information.

[0109] In some embodiments, the feature extraction layer can be implemented using a transformer, which may include three attention matrices (let it be represented as the first matrix W Q , the second matrix W K , the third matrix W V ), the three attention matrices can be used to increase the complexity of the target prediction model and enhance the expressiveness of the target prediction model so that the object extraction features can contain more feature information.

[0110] In some embodiments, the feature extraction layer is used to extract N object behavior features, specifically, each object behavior feature is multiplied by three attention matrices respectively to obtain three transition behavior features corresponding to each object behavior feature; it can be expressed as the following formula:

[0111] q i =H i *W Q

[0112] k i =H i *W K

[0113] v i =H i *W V

[0114] The above q i , k i 、v i represents the behavior characteristics of the i-th object (H i) respectively correspond to three transition behavior features, which represent the vector representation obtained by mapping the object behavior features based on different attention matrices; each object attribute information obtains the corresponding transition behavior feature in the above manner; the jth object behavior feature (H j ) are represented by the three transition behavior characteristics corresponding to q j , k j 、v j .

[0115] In some embodiments, the correlation between the i-th object behavior feature and the N object behavior features is determined based on the transition behavior feature. Specifically, it can be obtained by the following formula:

[0116]

[0117] Above I ij Indicates the correlation between the behavior characteristics of the i-th object and the behavior characteristics of the j-th object; j is an integer from 1 to N, and j can be equal to i; q i .k j Indicates q i With k j dot product; Indicates k j The square root of the number of dimensions.

[0118] In some embodiments, the feature information to be extracted from the N object behavior features under the i-th object behavior feature is determined based on the correlation degree. Specifically, it can be obtained by the following formula:

[0119] L ij =I ij *v j

[0120] The above L ij represents the feature information extracted from the jth object behavior feature under the i-th object behavior feature.

[0121] In some embodiments, the electronic device may determine the sum of feature information to be extracted from N object behavior features under the i-th object behavior feature as the object extraction feature corresponding to the i-th object behavior feature; or it may be that the sum of feature information is input into a fully connected layer connected to the feature extraction layer, and the output through the fully connected layer is determined as the final object extraction feature.

[0122] In one possible implementation, the number of feature extraction layers in the target prediction model can be multiple layers, each layer includes three attention matrices, the attention matrices of each layer are independent of each other, and each feature extraction layer can be followed by a fully connected layer. Multiple feature extractions can be performed on N object behavior features through the multiple feature extraction layers, and the output of the feature extraction layer of the previous layer is the input of the feature extraction layer of the next layer.

[0123] In some embodiments, the feature extraction layer corresponding to the resource behavior feature may be the same as or different from the feature extraction layer corresponding to the object behavior feature; and the fully connected layer connected to the corresponding feature extraction layer may be the same or different and independent of each other.

[0124] For example, Figure 6a-6c As shown, Figure 6a-6c An example diagram of a feature extraction scenario based on a target prediction model provided in an embodiment of the present application; wherein:

[0125] like Figure 6a As shown, assuming that the feature extraction layers corresponding to the object behavior features and the resource behavior features are the same, and the number of layers of the feature extraction layer is one, then the N object behavior features pass through the attention matrix 1-3 in the feature extraction layer and obtain the object extraction features of each object behavior feature through the fully connected layer, and the Q resource behavior features pass through the attention matrix 1-3 in the feature extraction layer and obtain the resource extraction features of each resource behavior feature through the fully connected layer;

[0126] Taking N object behavior features (let’s say H1, H2) as an example, based on the attention matrix 1-3 (W Q , W K , W V ) determine the transition behavior characteristics (q1, k1, v1) of H1 and the transition behavior characteristics (q2, k2, v2) of H2; determine the correlation between H1 and H2 based on the transition behavior characteristics 11 and the degree of correlation with H2 12 , and determine the degree of correlation between H2 and H1 21 and the degree of correlation with H2 22 According to I 11 Determine the feature information L extracted from H1 under H1 11 and according to I 12 Determine the feature information L extracted from H2 12 , according to I 21 Determine the feature information L extracted from H1 under H2 21 and according to I 22 Determine the feature information L extracted from H2 22 ; L 11 and L 12The sum is determined by the output of the fully connected layer as the object extraction feature Z1 of H1, and L 21 and L 22 The sum is determined through the output of the fully connected layer as the object extraction feature Z2 of H2;

[0127] like Figure 6b As shown, assuming that the feature extraction layers corresponding to the object behavior feature and the resource behavior feature are different, and the number of feature extraction layers is one, then the N object behavior features pass through the attention matrix 1-3 in the feature extraction layer 1 and obtain the object extraction feature of each object behavior feature through the fully connected layer, and the Q resource behavior features pass through the attention matrix 4-6 in the feature extraction layer 2 and obtain the resource extraction feature of each resource behavior feature through the fully connected layer;

[0128] like Figure 6c As shown, assuming that the feature extraction layers corresponding to the object behavior feature and the resource behavior feature are different, and the number of feature extraction layers is 3, then the N object behavior features pass through the attention matrix 1-9 in the feature extraction layers 1.1-1.3 and obtain the object extraction feature of each object behavior feature through the fully connected layers 1-3, and the Q resource behavior features pass through the attention matrix 10-18 in the feature extraction layers 2.1-2.3 and obtain the resource extraction feature of each resource behavior feature through the fully connected layers 1-3;

[0129] Take N object behavior features (let’s represent them as H1 0 、H2 0 ) Taking the feature extraction layer 1.1-1.2 as an example, H1 and H2 extract features through the attention matrix 1-3 in the feature extraction layer 1.1 and output the initial object extraction features through the fully connected layer 1 (let it be represented as H1 1 、H2 1 ), and change the H1 1 、H2 1 The input feature extraction layer 1.2 continues to perform feature extraction, thereby obtaining the object extraction features of the final output.

[0130] S405: Determine the target object behavior feature according to the influencing factor of the object behavior feature and the object extraction feature, and determine the target resource behavior feature according to the influencing factor of the resource behavior feature and the resource extraction feature.

[0131] In some embodiments, the electronic device may determine the target object behavior feature corresponding to the ith object behavior feature based on the influencing factor of the ith object behavior feature and the object extraction feature of the ith object behavior feature. Specifically, the product of the influencing factor and the object behavior feature may be used as the target object behavior feature. Similarly, the electronic device may determine the target resource behavior feature corresponding to the qth resource behavior feature based on the influencing factor of the qth resource behavior feature and the resource extraction feature of the qth resource behavior feature. Specifically, the product of the influencing factor and the resource behavior feature may be used as the target resource behavior feature.

[0132] S406: Determine the target probability of the target object responding to the recommended resource according to the object attribute characteristics, resource attribute characteristics, target object behavior characteristics, and target resource behavior characteristics.

[0133] In some embodiments, the electronic device may obtain the target probability through a target prediction model. The target prediction model may include a probability prediction network corresponding to each resource to be recommended. In the target prediction model, the corresponding probability prediction network may be determined according to the resource type identifier carried in the recommendation information set, and the target probability may be determined by the probability prediction network according to the object attribute characteristics, resource attribute characteristics, target object behavior characteristics, and target resource behavior characteristics.

[0134] S407: Push the recommended resources to the target terminal of the target object according to the target probability.

[0135] In some embodiments, the electronic device can determine the recommended resources that the target object is most likely to be interested in based on the target probability, thereby improving the effective recommendation rate. Therefore, the recommended resources are pushed according to the target probability. The push method is not limited here. Alternatively, after predicting the target probabilities of a large number of target objects for different recommended resources, the recommended resources can also be divided into groups based on the target probability. For example, the target object can be divided into multiple target groups, and the recommended resources suitable for recommendation in the different target groups are determined according to the target probability. When there is a new target object to be recommended later, there is no need to predict the corresponding target probability, but the target group to which the new target object belongs can be determined, and the recommended resources suitable for recommendation for the target group can be pushed to the target group to which the new target object belongs.

[0136] For example, Figure 7 As shown, Figure 7 A schematic diagram of a push scenario of resources to be recommended based on a target prediction model provided in an embodiment of the present application; taking the recommendation task applied to a live broadcast room as an example, the recommendation principle of other recommendation tasks is the same as the recommendation task for the live broadcast room; wherein:

[0137] When the target application in the user terminal displays a product scenario that can recommend a live broadcast room, the electronic device obtains multiple live broadcast rooms (i1, i2, i3, ..., iN) that are currently broadcasting live, and respectively forms a prediction pair ([u, i1], [u, i2], [u, i3], ..., [u, iN]) with the target user (u) and each live broadcast room, and extracts user attribute information of the target user, user behavior information with respect to the live broadcast room, live broadcast room attribute information and live broadcast room behavior information of each live broadcast room from an attribute storage platform (such as a database) based on the multiple prediction pairs, so as to store the aforementioned multiple information and the type identifier (s1, ..., s N) form information pairs ([u,i1,s1], [u,i2,s2], [u,i3,s3], ..., [u,iN,sN]), and input each information pair into the target prediction model in turn; in the target prediction model, generate a feature pair corresponding to each information pair according to the type identifier and the feature generation method defined by the live broadcast room recommendation task, and obtain the target probability under each prediction pair according to the feature pair corresponding to each prediction pair; each live broadcast room can be sorted according to the target probability under each prediction pair to obtain the sorted live broadcast rooms, and the sorted live broadcast rooms or the target number of sorted live broadcast rooms can be pushed to the user terminal of the target user in turn.

[0138] In an embodiment of the present application, a recommendation information set can be obtained, a recommendation feature set can be generated based on the recommendation information set, an influencing factor of an object behavior feature and an influencing factor of a resource behavior feature can be determined based on an object attribute feature and a resource attribute feature, a feature extraction can be performed on the object behavior feature to obtain an object extraction feature of the object behavior feature, and a feature extraction can be performed on the resource behavior feature to obtain a resource extraction feature of the resource behavior feature, a target object behavior feature can be determined based on the influencing factor of the object behavior feature and the object extraction feature, and a target resource behavior feature can be determined based on the influencing factor of the resource behavior feature and the resource extraction feature, a target probability that the target object responds to the recommended behavior of the recommended resource can be determined based on the object attribute feature, the resource attribute feature, the target object behavior feature and the target resource behavior feature, and the recommended resource can be pushed to the target terminal of the target object according to the target probability. Through the above method, not only can the target probability be determined by combining more feature information, but also the object behavior feature and the resource behavior feature can be feature selected by the influencing factor to determine the target behavior feature that is more suitable in the prediction process, so that the recommendation prediction efficiency and accuracy of the recommended resource can be improved, which is conducive to improving the effect of subsequent recommendation based on the target probability.

[0139] See also Figure 8 , Figure 8 The following is a flow chart of a data processing method provided in an embodiment of the present application, which can be executed by the electronic device mentioned above. Figure 8As shown, the process of the data processing method in the embodiment of the present application may include the following:

[0140] S801. Obtain M sample information sets.

[0141] A sample information set includes sample object attribute information and sample object behavior information of a sample object, and sample resource attribute information and sample resource behavior information of a sample recommended resource.

[0142] In some embodiments, to achieve joint recommendation of cross-domain recommendation tasks, each sample information set has a probability label and a resource type label. M sample information sets have R resource type labels, and one resource type label corresponds to one recommendation task. M is a positive integer greater than 1, and R is a positive integer greater than 1 and less than or equal to M. The sample objects and sample recommended resources in the M sample information sets may be different.

[0143] In some embodiments, the above-mentioned target probability can be generated by a target prediction model. Therefore, the electronic device can use the M sample information sets as a round of sample data for batch training of the initial prediction model, and obtain the final target prediction model based on multiple rounds of training of M sample information sets until convergence. The M in each round can be the same or different. In each round of M sample information sets, the number of sample information sets of the same resource type label can be specified or randomly generated. The process and principle of each round of training are the same, and the training of the M sample information sets in the kth round is used as an example for explanation. Among them, the relevant description of the sample data contained in the M sample information sets can refer to the description of the data used in the application process described in the above embodiment.

[0144] S802: Input each sample information set into an initial prediction model, and generate a sample probability corresponding to each sample information set based on the initial prediction model.

[0145] In a possible implementation, any sample information set among the M sample information sets is represented as the jth sample information set, where j is a positive integer less than or equal to M. The electronic device generating the sample probability corresponding to each sample information set based on the initial prediction model may specifically be: generating a sample feature set corresponding to the jth sample information set based on the initial prediction model, the sample feature set including sample object attribute features of the sample object attribute information, sample object behavior features of the sample object behavior information, sample resource attribute features of the sample resource attribute information, and sample resource behavior features of the sample resource behavior features in the jth sample information set; generating a sample influence factor of the sample object behavior feature and a sample influence factor of the sample resource behavior feature according to the sample object attribute features and the sample resource attribute features, determining a target sample object behavior feature according to the sample influence factor of the sample object behavior feature and the sample object behavior feature, and determining a target sample resource behavior feature according to the sample influence factor of the sample resource behavior feature and the sample resource behavior feature, and determining the sample probability corresponding to the jth sample information set according to the object attribute features, the resource attribute features, the target sample object behavior features, and the target sample resource behavior features. Among them, the sample feature set can also include sample features generated for resource type labels. The generation method can be the same as the generation method of other sample features. When subsequently inputting the initial probability prediction network, the sample features generated for the resource type labels and the connected sample feature set can also be input together to predict the sample probability.

[0146] In some embodiments, the initial prediction model may include a feature generation layer to be trained, and the model parameters in the feature generation layer may include an embedded feature matrix, which can be obtained by model training. The electronic device may input each sample information set into the initial prediction model in sequence, and perform feature encoding on each sample information set based on the feature generation layer in the initial prediction model and according to the matching encoding rules to obtain a coded feature vector, and obtain a corresponding sample feature set according to the coded feature vector and the embedded feature matrix; the specific process and principle of generating the sample feature set can refer to the relevant description of the above embodiment. Among them, the encoding rules matched by different sample information sets may be different. When encoding, the features may be encoded according to the encoding rules matched by each sample information set, or according to a unified encoding rule. For example, the features are encoded according to the encoding rules matched in the recommendation task for the live broadcast room. At this time, the recommendation task corresponding to the unified encoding rule can also be expressed as the main recommendation task (main domain) set in the cross-domain recommendation task, that is, it can be understood that the training is focused on the main recommendation task. Further, after the initial prediction model is trained with the main recommendation task, other recommendation tasks can be used as the main recommendation tasks to continue training the model to obtain the final model.

[0147] In some embodiments, during the training process, feature alignment can be performed on common features in different recommendation tasks, and the embedded feature matrices corresponding to the aligned features are the same. By aligning common features, more feature information in cross-domain recommendation tasks during training can be transferred to common features, so that feature selection of behavior information under different recommendation tasks can be achieved through common features. Compared with selecting sample data obtained based on the intersection of users of each recommendation task for training, the technical solution of the present application is directly based on all sample data in each recommendation task for training, which can improve training efficiency; by aligning the common features, a correlation relationship between multiple recommendation tasks can be established, without sample screening, and more feature information can be learned to enhance the model effect. And, in the application process, if the user's behavior information in the current recommendation task is insufficient, since the common features also transmit feature information under other recommendation tasks, it can be achieved to indirectly use the behavior information in other recommendation tasks to make up for the information of the current recommendation task, which can well alleviate the cold start problem and further improve the recommendation effect for new users in the current recommendation task.

[0148] In some embodiments, the initial prediction model may include an influence factor prediction network to be trained. The electronic device may generate the sample influence factor of the sample object behavior feature and the sample influence factor of the sample resource behavior feature according to the sample object attribute feature and the sample resource attribute feature by splicing the sample object attribute feature and the sample resource attribute feature and inputting the influence factor prediction network to obtain the sample influence factor; or may also perform feature fusion on the sample object attribute feature and the sample resource attribute feature to obtain a sample fusion attribute feature, and input the sample fusion attribute feature into the influence factor prediction network to obtain the sample influence factor. The specific process of the feature fusion can refer to the relevant description of the above embodiment.

[0149] In some embodiments, the process and principle of determining the target sample object behavior characteristics and determining the target sample resource behavior characteristics are the same, and the determination of the target sample object behavior characteristics is used as an example for explanation. Determining the target sample object behavior characteristics based on the sample influence factor of the sample object behavior characteristics and the sample object behavior characteristics can be to use the product of the sample influence factor and the sample object behavior characteristics as the target sample object behavior characteristics; or it can also be that the initial prediction model includes a feature extraction layer (or also includes a fully connected layer), and the sample object behavior characteristics are extracted according to the feature extraction layer (or also includes a fully connected layer) to obtain the sample extraction behavior characteristics, and the product of the sample influence factor and the sample extraction behavior characteristics is used as the target sample object behavior characteristics. The specific process of performing feature extraction can refer to the relevant description of the above embodiment.

[0150] In some embodiments, the initial prediction model may include R initial probability prediction networks, and one resource type label corresponds to one initial probability prediction network. The electronic device determines the sample probability corresponding to the jth sample information set based on the object attribute characteristics, resource attribute characteristics, target sample object behavior characteristics, and target sample resource behavior characteristics. In the jth initial probability prediction network, the sample probability corresponding to the jth sample information set is generated based on the object attribute characteristics, resource attribute characteristics, target sample object behavior characteristics, and target sample resource behavior characteristics. Among them, the jth initial probability prediction network is the initial probability prediction network corresponding to the jth resource type label in the R initial probability prediction networks, and the jth resource type label is the resource type label of the jth sample information set in the R resource type labels. That is, the sample information set is predicted using the corresponding initial probability prediction network to obtain the corresponding sample probability.

[0151] S803: Determine a target loss function for M sample information sets according to the number of each resource type label.

[0152] In one possible implementation, when jointly training cross-domain recommendation tasks, the main recommendation task can be determined from the cross-domain recommendation tasks, which can mean focusing on the main recommendation task for training or balancing the training process of the main recommendation task, so that the target prediction model can better learn the feature information of the main recommendation task itself and the feature information contributed by other recommendation tasks. The main recommendation task can be any recommendation task set by relevant business personnel, or it can be a recommendation task that meets preset conditions. The prediction condition can, for example, refer to the recommendation task with the least number of samples; for example, the number of sample information sets of recommendation task 1 is 1000, and the number of sample information sets of recommendation task 2 is 100, so the target prediction model has less information learned based on recommendation task 2. In order to balance the training process of the two recommendation tasks, recommendation task 2 can be set as the main recommendation task. Furthermore, during the model training process, the main recommendation task can also be replaced in turn for training, so that the target prediction model can have a good recommendation effect for all recommendation tasks.

[0153] In some embodiments, R resource type labels are associated with R initial loss functions; determining the target loss function for M sample information sets based on the number of each resource type label can specifically be, obtaining a target probability prediction network from the R initial probability prediction networks, determining the training weight of the target initial loss function associated with the resource type label corresponding to the target probability prediction network based on the number of resource type labels corresponding to the target probability prediction network, and determining the target loss function based on the training weight of the target initial loss function and the R initial loss functions. The target probability prediction network is the initial probability prediction network corresponding to the main recommendation task. This method can achieve dynamic adjustment of the target loss function according to the number of resource type labels during each round of model training, so that the determination of the target loss function is more flexible and more adapted to specific training scenarios.

[0154] In some embodiments, the objective loss function Loss can be expressed as the following formula:

[0155] Loss=Loss1+......+f(batch,curr)Loss R

[0156] Among them, Loss1, ..., Loss R Represents R initial loss functions, Loss R represents the target initial loss function corresponding to the main recommendation task, f(x1,x2) represents the determination function of the training weight, batch represents the number of M, and curr represents the number of resource type labels corresponding to the main recommendation task in the M sample information sets.

[0157] In some embodiments, the training weight determination function f(x1, x2) can be expressed as:

[0158] f(x1,x2)=log(x1 / x2)

[0159] Or, it can also be expressed as:

[0160] It is understandable that in different recommendation tasks, the number of samples usually varies greatly. If the initial loss functions of multiple recommendation tasks are directly combined to train the model, the feature information in the recommendation tasks with fewer samples may be diluted, which is not conducive to the overall training of the model. Therefore, the recommendation task with fewer samples can be used as the main recommendation task, and the initial loss function corresponding to the main recommendation task can be weighted based on the above method to ensure the utilization rate of the feature information of the main recommendation task.

[0161] In one possible implementation, if the number of samples for each recommended task is different, the training effect of the target prediction model for each recommended task is different. Therefore, in order to balance the training effect of each recommended task, the target loss function for the M sample information sets is determined according to the number of each resource type label. Specifically, the training weight of each initial loss function is determined according to the number of each resource type label, and the R initial loss functions are weighted and summed according to the training weight of each initial loss function to obtain the target loss function. The target loss function Loss can be expressed as the following formula:

[0162] Loss=f(batch,curr1)Loss1+......+f(batch,curr R )Loss R

[0163] Among them, Loss1, ..., Loss R represents R initial loss functions, f(x1,x2) represents the determination function of training weights, batch represents the number of M, curr1, ..., curr R Indicates the number of resource type labels corresponding to each recommendation task in the M sample information sets.

[0164] In some embodiments, the training weight determination function f(x1, x2) can be expressed as the following formula:

[0165] f(x1,x2)=log(x1 / x2)

[0166] S804: Train the initial prediction model according to the target loss function, the sample probability corresponding to each sample information set, and the corresponding probability label to obtain a target prediction model.

[0167] In some embodiments, the loss value for M sample information sets can be obtained based on the target loss function, the sample probability corresponding to each sample information set, and the corresponding probability label, and the loss value can be used to modify the model parameters of the initial prediction model to train the initial prediction model to obtain the target prediction model.

[0168] For example, Figure 9a As shown, Figure 9a A schematic diagram of a scenario of model training provided in an embodiment of the present application; the initial prediction model may include a feature generation layer, an impact factor prediction network, and a probability prediction layer, and the probability prediction layer includes multiple initial probability prediction networks; the impact factor prediction network and the initial probability prediction network may both be composed of one or more fully connected layers, and the number of fully connected layers of different prediction networks may be different, which can be set by relevant business personnel according to experience; wherein:

[0169] Assume that the cross-domain recommendation tasks are recommendation task A and recommendation task B, and recommendation task A is the main recommendation task; when M sample information sets are obtained (set 1, set 2, and set 3), determine the number of resource type tags corresponding to recommendation task A in the M sample information sets; assume that the resource type tags of set 1 and set 2 correspond to recommendation task A, and the resource type tags of set 3 correspond to recommendation task B;

[0170] Generate a sample feature set corresponding to each sample information set in the feature generation layer, the sample feature set including a first sample feature corresponding to a common feature (such as an object attribute feature and a resource attribute feature) and a second sample feature corresponding to a non-common feature (such as an object behavior feature and a resource behavior feature);

[0171] In the impact factor prediction network (composed of a fully connected layer), feature selection is performed on the second sample feature according to the first sample feature to obtain the impact factor of each second sample feature, and the target second sample feature is determined according to the impact factor and the second sample feature;

[0172] In the initial probability prediction network (composed of a fully connected layer) corresponding to each of the M sample information sets, the sample probability corresponding to each sample information set is determined according to the first sample feature and the target second sample feature; the set 1 and the set 2 correspond to the initial probability prediction network 1, and the set 3 corresponds to the initial probability prediction network 2;

[0173] According to the number of resource type labels corresponding to the recommended task A, the training weight of the initial loss function for the recommended task A is determined (which can be expressed as f(3,2)), and the target loss function is determined according to the training weight and the initial loss function of each recommended task A. The parameters of the initial prediction model are modified according to the target loss function to obtain the target prediction model.

[0174] It is understandable that the model structure of the above initial prediction model is only an example, and other forms are also possible. For example, the initial prediction network may include a feature extraction layer, through which the object behavior feature (resource behavior feature) is extracted to obtain the corresponding object extraction feature (resource extraction feature), and the target object behavior feature (target resource behavior feature) input into the initial probability prediction network is determined through the object extraction feature and the influencing factor, such as Figure 9b As shown, Figure 9b A schematic diagram of a prediction model provided in an embodiment of the present application.

[0175] Through the above process, model prediction of cross-domain recommendation tasks can be realized, and feature adaptation of cross-domain training can be achieved. The cross-domain recommendation task can use the feature information learned from the current domain and other domains to jointly act on the recommendation prediction of the current domain. In addition, when the model is trained in the main recommendation task mode, the training effect of the main recommendation task can be strengthened, and the feature information contributed by other domains can be efficiently transferred to the main domain through joint training of multiple sample data and dynamic adjustment of the loss function.

[0176] After a large number of tests on the target prediction model obtained through the above training, it was found that the prediction accuracy and efficiency of the target prediction model were greatly improved compared with the existing technology, and through this method, each recommendation task in the target application (assuming that it includes recommendation task 1 for short videos, recommendation task 2 for karaoke rooms, and recommendation task 3 for live broadcast rooms) has a significant improvement in viewing time. See the following Table 1:

[0177] Dataset Exposure Watch time Short Video 1618.44 million 323.68 million singing room 597.62 million 119.52 million live streaming 74.51 million 14.9 million

[0178] Table 1

[0179] Among them, the target prediction model was tested using the data set, and it was found that the push exposure for short videos was about 1618.44 million times, and the effective viewing time was about 323.68 million times; the push exposure for karaoke rooms was about 597.62 million times, and the effective viewing time was 119.52 million times; the push exposure for live broadcasts was about 74.51 million times, and the effective viewing time was 14.90 million times.

[0180] By comparing with the existing prediction model based on directly integrating the sample data of the intersection users, it is found that the model evaluation index AUC (area under the curve) of the target prediction model proposed by the technical solution of this application is also greatly improved. The larger the AUC, the better the training effect of the model and the higher the prediction accuracy. See Table 2 below:

[0181] Model AUC Existing prediction models 0.8259 Target prediction model 0.8346

[0182] Table 2

[0183] Furthermore, 10% of online users were selected for comparison, and it was found that the average viewing time and total viewing time obtained when the target prediction model proposed in the technical solution of this application was applied were better than those obtained when the above-mentioned existing prediction model was applied. See Table 3 below:

[0184]

[0185] Table 3

[0186] Among them, the average viewing time per person when the target prediction model is applied is about 2.02% higher than the average viewing time per person when the existing prediction model is applied; the total viewing time when the target prediction model is applied is about 2.36% higher than the total viewing time when the existing prediction model is applied.

[0187] Furthermore, when making recommendation predictions for new users whose current domain behavior is insufficient, the target prediction model proposed by the technical solution of this application is also superior to the above-mentioned existing prediction models in terms of average viewing time and total viewing time. See Table 4 below:

[0188]

[0189] Table 4

[0190] Among them, the average viewing time of new users when the target prediction model is applied is about 3.59% higher than the average viewing time of new users when the existing prediction model is applied; the total viewing time of new users when the target prediction model is applied is about 3.48% higher than the total viewing time of new users when the existing prediction model is applied.

[0191] In an embodiment of the present application, M sample information sets can be obtained, each sample information set is input into an initial prediction model, a sample probability corresponding to each sample information set is generated based on the initial prediction model, a target loss function for the M sample information sets is determined according to the number of labels of each resource type, and the initial prediction model is trained according to the target loss function, the sample probability corresponding to each sample information set, and the corresponding probability label to obtain a target prediction model. Through the above method, joint training using multiple sample recommendation resources containing different types can enable the target prediction model to jointly recommend multiple resources to be recommended, and the target loss function can be dynamically adjusted during training, thereby improving the training effect of the model.

[0192] See also Fig.10 , Fig.10 This is a structural diagram of a data processing device provided in the present application; the data processing device can be used as a computer program (including program code) running in an electronic device, for example, the data processing device can be an application program in the electronic device (such as a program that can perform data processing). It should be noted that Fig.10 The data processing device shown is used to execute the application Figure 2 , Figure 4 and Figure 8 Some or all steps in the method of the embodiment shown. The data processing device 1000 may include: an acquisition module 1001, a processing module 1002, and a determination module 1003. Among them:

[0193] The acquisition module 1001 is used to acquire a recommendation information set; the recommendation information set includes object attribute information and object behavior information of the target object, resource attribute information and resource behavior information of the resource to be recommended;

[0194] Processing module 1002, for generating a recommended feature set according to the recommended information set; the recommended feature set includes object attribute features of object attribute information, object behavior features of object behavior information, resource attribute features of resource attribute information, and resource behavior features of resource behavior information;

[0195] Determination module 1003, used to determine the influencing factors of the object behavior characteristics and the influencing factors of the resource behavior characteristics according to the object attribute characteristics and the resource attribute characteristics;

[0196] The determination module 1003 is further used to determine the target object behavior characteristics according to the influencing factors of the object behavior characteristics and the object behavior characteristics, and to determine the target resource behavior characteristics according to the influencing factors of the resource behavior characteristics and the resource behavior characteristics;

[0197] The determination module 1003 is further used to determine the target probability of the target object responding to the recommended resource based on the object attribute characteristics, resource attribute characteristics, target object behavior characteristics and target resource behavior characteristics, and push the recommended resource to the target object's object terminal according to the target probability.

[0198] In a possible implementation, the target probability is generated by a target prediction model; the processing module 1002 is further configured to:

[0199] Obtain M sample information sets; one sample information set includes sample object attribute information and sample object behavior information of a sample object, and sample resource attribute information and sample resource behavior information of a sample recommended resource; each sample information set has a probability label and a resource type label, and the M sample information sets have R resource type labels, where M is a positive integer greater than 1, and R is a positive integer greater than 1 and less than or equal to M;

[0200] Input each sample information set into the initial prediction model, and generate the sample probability corresponding to each sample information set based on the initial prediction model; any sample probability represents the probability of the sample object in the corresponding sample information set responding to the sample recommendation resource in the sample information set;

[0201] Determine the target loss function for the M sample information sets according to the number of labels of each resource type;

[0202] The initial prediction model is trained according to the target loss function, the sample probability corresponding to each sample information set and the corresponding probability label to obtain the target prediction model.

[0203] In a possible implementation, when the processing module 1002 is used to generate the sample probability corresponding to each sample information set based on the initial prediction model, it is specifically used to:

[0204] Generate a sample feature set corresponding to the j-th sample information set based on the initial prediction model; the sample feature set includes sample object attribute features of the sample object attribute information, sample object behavior features of the sample object behavior information, sample resource attribute features of the sample resource attribute information, and sample resource behavior features of the sample resource behavior information in the j-th sample information set; the j-th sample information set is any sample information set among the M sample information sets, and j is a positive integer less than or equal to M;

[0205] Generate a sample impact factor of the sample object behavior feature and a sample impact factor of the sample resource behavior feature according to the sample object attribute feature and the sample resource attribute feature;

[0206] Determine the target sample object behavior characteristics according to the sample influence factor of the sample object behavior characteristics, and determine the target sample resource behavior characteristics according to the sample influence factor of the sample resource behavior characteristics;

[0207] The sample probability corresponding to the j-th sample information set is determined according to the object attribute characteristics, the resource attribute characteristics, the target sample object behavior characteristics and the target sample resource behavior characteristics.

[0208] In a possible implementation, the initial prediction model includes R initial probability prediction networks, and one resource type label corresponds to one initial probability prediction network;

[0209] When the processing module 1002 is used to determine the sample probability corresponding to the j-th sample information set according to the object attribute characteristics, the resource attribute characteristics, the target sample object behavior characteristics and the target sample resource behavior characteristics, it is specifically used to:

[0210] In the jth initial probability prediction network, the sample probability corresponding to the jth sample information set is generated according to the object attribute characteristics, resource attribute characteristics, target sample object behavior characteristics and target sample resource behavior characteristics;

[0211] Among them, the jth initial probability prediction network is the initial probability prediction network corresponding to the jth resource type label in the R initial probability prediction networks, and the jth resource type label is the resource type label of the jth sample information set in the R resource type labels.

[0212] In a possible implementation, the initial prediction model includes R initial probability prediction networks, and the R resource type labels are associated with R initial loss functions; when the processing module 1002 is used to determine the target loss function for the M sample information sets according to the number of each resource type label, it is specifically used to:

[0213] Obtain a target probability prediction network from R initial probability prediction networks;

[0214] Determine the training weight of the target initial loss function corresponding to the resource type label corresponding to the target probability prediction network according to the number of resource type labels corresponding to the target probability prediction network;

[0215] The target loss function is determined based on the training weight of the target initial loss function and the R initial loss functions.

[0216] In a possible implementation, R resource type tags are associated with R initial loss functions; when the processing module 1002 is used to determine the target loss function for the M sample information sets according to the number of each resource type tag, it is specifically used to:

[0217] Determine the training weight of each initial loss function based on the number of labels for each resource type;

[0218] The target loss function is obtained by weighted summing the R initial loss functions according to the training weight of each initial loss function.

[0219] In a possible implementation, when the determination module 1003 is used to determine the influencing factors of the object behavior characteristics and the influencing factors of the resource behavior characteristics according to the object attribute characteristics and the resource attribute characteristics, it is specifically used to:

[0220] Perform feature fusion on the object attribute features and the resource attribute features to obtain fused attribute features;

[0221] The influencing factors of the object behavior characteristics and the influencing factors of the resource behavior characteristics are generated according to the fused attribute characteristics.

[0222] In a possible implementation, the number of the object behavior features is N, where N is a positive integer greater than 1, and any one of the N object behavior features is represented as the i-th object behavior feature, where i is a positive integer less than or equal to N;

[0223] When the determination module 1003 is used to determine the target object behavior feature according to the influencing factor of the object behavior feature and the object behavior feature, it is specifically used to:

[0224] Perform feature extraction on the i-th object behavior feature according to the N object behavior features to obtain an object extraction feature of the i-th object behavior feature;

[0225] The target object behavior feature corresponding to the ith object behavior feature is determined according to the object extraction feature of the ith object behavior feature and the influencing factor of the ith object behavior feature.

[0226] According to one embodiment of the present application, Fig.10 The various units in the data processing device shown can be combined into one or more other units separately or in whole, or one (or more) of the units can be further divided into multiple smaller units in terms of function, which can achieve the same operation without affecting the realization of the technical effects of the embodiments of the present application. The above units are divided based on logical functions. In practical applications, the functions of one unit can also be realized by multiple units, or the functions of multiple units can be realized by one unit.

[0227] In other embodiments of the present application, the data processing device may also include other units. In practical applications, these functions may also be implemented with the assistance of other units, and may be implemented by multiple units in collaboration. According to another embodiment of the present application, the data processing device may be executed by running on a general computing device such as a computer including a central processing unit (CPU), a random access storage medium (RAM), a read-only storage medium (ROM), and other processing elements and storage elements. Figure 2 , Figure 4 and Figure 8 A computer program (including program code) for each step involved in the corresponding method shown in FIG. Fig.10 The data processing device shown in the embodiment of the present application is implemented by the data processing method. The computer program can be recorded on a computer-readable recording medium, for example, and loaded into the above-mentioned electronic device through the computer-readable recording medium and run therein.

[0228] In the embodiment of the present application, the acquisition module acquires the recommendation information set; the processing module generates the recommendation feature set according to the recommendation information set; the recommendation feature set includes the object attribute feature of the object attribute information, the object behavior feature of the object behavior information, the resource attribute feature of the resource attribute information, and the resource behavior feature of the resource behavior information; the determination module determines the influencing factor of the object behavior feature and the influencing factor of the resource behavior feature according to the object attribute feature and the resource attribute feature; the determination module determines the target object behavior feature according to the influencing factor of the object behavior feature and the object behavior feature, and determines the target resource behavior feature according to the influencing factor of the resource behavior feature and the resource behavior feature; the determination module determines the target probability that the target object responds to the recommended behavior of the recommended resource according to the object attribute feature, the resource attribute feature, the target object behavior feature and the target resource behavior feature, and pushes the recommended resource to the object terminal of the target object according to the target probability. Through the above device, not only can the target probability be determined in combination with more feature information, but also the object behavior feature and the resource behavior feature can be selected by the influencing factor to determine the target behavior feature that is more suitable in the prediction process, so that the recommendation prediction efficiency and accuracy of the recommended resource can be improved, which is conducive to improving the effect of subsequent recommendation based on the target probability.

[0229] Each functional module in each embodiment of the present application can be integrated into one module, or each module can exist physically separately, or two or more modules can be integrated into one module. The above integrated modules can be implemented in the form of hardware or software functional modules, which is not limited in the present application.

[0230] See also Fig.11 , Fig.11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Fig.11 As shown, the electronic device 1100 includes: a processor 1101, a communication interface 1102 and a computer-readable storage medium 1103. The processor 1101, the communication interface 1102 and the computer-readable storage medium 1103 can be connected via a bus or other means. The communication interface 1102 is used to receive and send data. The computer-readable storage medium 1103 can be stored in the memory of the electronic device, and the computer-readable storage medium 1103 is used to store a computer program, and the computer program includes program instructions. The processor 1101 is used to execute the program instructions stored in the computer-readable storage medium 1103. The processor 1101 (or CPU (Central Processing Unit)) is the computing core and control core of the electronic device, which is suitable for implementing one or more instructions, and is specifically suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding function.

[0231] The embodiment of the present application also provides a computer-readable storage medium (Memory), which is a memory device in an electronic device for storing programs and data. It is understandable that the computer-readable storage medium here can include both built-in storage media in electronic devices and, of course, extended storage media supported by electronic devices. The computer-readable storage medium provides a storage space that stores the processing system of the electronic device. In addition, one or more instructions suitable for being loaded and executed by the processor 1101 are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory, or a non-volatile memory, such as at least one disk storage; optionally, it can also be at least one computer-readable storage medium located away from the aforementioned processor.

[0232] In one embodiment, the computer-readable storage medium stores one or more instructions; the processor 1101 loads and executes the one or more instructions stored in the computer-readable storage medium to implement the corresponding steps in the above document processing method embodiment; in a specific implementation, the one or more instructions in the computer-readable storage medium are loaded by the processor 1101 and the following steps are executed:

[0233] Acquire a recommendation information set; the recommendation information set includes object attribute information and object behavior information of the target object, resource attribute information and resource behavior information of the resource to be recommended;

[0234] Generate a recommended feature set according to the recommended information set; the recommended feature set includes object attribute features of object attribute information, object behavior features of object behavior information, resource attribute features of resource attribute information, and resource behavior features of resource behavior information;

[0235] Determine the influencing factors of the object behavior characteristics and the influencing factors of the resource behavior characteristics according to the object attribute characteristics and the resource attribute characteristics;

[0236] Determine the target object behavior characteristics according to the influencing factors of the object behavior characteristics and the object behavior characteristics, and determine the target resource behavior characteristics according to the influencing factors of the resource behavior characteristics and the resource behavior characteristics;

[0237] The target probability of the target object responding to the recommended resource is determined according to the object attribute characteristics, resource attribute characteristics, target object behavior characteristics and target resource behavior characteristics, and the recommended resource is pushed to the target object's object terminal according to the target probability.

[0238] In a possible implementation, the target probability is generated by a target prediction model; the processor 1101 is further configured to:

[0239] Obtain M sample information sets; one sample information set includes sample object attribute information and sample object behavior information of a sample object, and sample resource attribute information and sample resource behavior information of a sample recommended resource; each sample information set has a probability label and a resource type label, and the M sample information sets have R resource type labels, where M is a positive integer greater than 1, and R is a positive integer greater than 1 and less than or equal to M;

[0240] Input each sample information set into the initial prediction model, and generate the sample probability corresponding to each sample information set based on the initial prediction model; any sample probability represents the probability of the sample object in the corresponding sample information set responding to the sample recommendation resource in the sample information set;

[0241] Determine the target loss function for the M sample information sets according to the number of labels of each resource type;

[0242] The initial prediction model is trained according to the target loss function, the sample probability corresponding to each sample information set and the corresponding probability label to obtain the target prediction model.

[0243] In a possible implementation, when the processor 1101 is used to generate the sample probability corresponding to each sample information set based on the initial prediction model, it is specifically used to:

[0244] Generate a sample feature set corresponding to the j-th sample information set based on the initial prediction model; the sample feature set includes sample object attribute features of the sample object attribute information, sample object behavior features of the sample object behavior information, sample resource attribute features of the sample resource attribute information, and sample resource behavior features of the sample resource behavior information in the j-th sample information set; the j-th sample information set is any sample information set among the M sample information sets, and j is a positive integer less than or equal to M;

[0245] Generate a sample impact factor of the sample object behavior feature and a sample impact factor of the sample resource behavior feature according to the sample object attribute feature and the sample resource attribute feature;

[0246] Determine the target sample object behavior characteristics according to the sample influence factor of the sample object behavior characteristics, and determine the target sample resource behavior characteristics according to the sample influence factor of the sample resource behavior characteristics;

[0247] The sample probability corresponding to the j-th sample information set is determined according to the object attribute characteristics, the resource attribute characteristics, the target sample object behavior characteristics and the target sample resource behavior characteristics.

[0248] In a possible implementation, the initial prediction model includes R initial probability prediction networks, and one resource type label corresponds to one initial probability prediction network;

[0249] When the processor 1101 is used to determine the sample probability corresponding to the j-th sample information set according to the object attribute characteristics, the resource attribute characteristics, the target sample object behavior characteristics, and the target sample resource behavior characteristics, it is specifically used to:

[0250] In the jth initial probability prediction network, the sample probability corresponding to the jth sample information set is generated according to the object attribute characteristics, resource attribute characteristics, target sample object behavior characteristics and target sample resource behavior characteristics;

[0251] Among them, the jth initial probability prediction network is the initial probability prediction network corresponding to the jth resource type label in the R initial probability prediction networks, and the jth resource type label is the resource type label of the jth sample information set in the R resource type labels.

[0252] In a possible implementation, the initial prediction model includes R initial probability prediction networks, and the R resource type labels are associated with R initial loss functions; when the processor 1101 is used to determine the target loss function for the M sample information sets according to the number of each resource type label, it is specifically used to:

[0253] Obtain a target probability prediction network from R initial probability prediction networks;

[0254] Determine the training weight of the target initial loss function corresponding to the resource type label corresponding to the target probability prediction network according to the number of resource type labels corresponding to the target probability prediction network;

[0255] The target loss function is determined based on the training weight of the target initial loss function and the R initial loss functions.

[0256] In a possible implementation, R resource type labels are associated with R initial loss functions; when the processor 1101 is used to determine the target loss function for the M sample information sets according to the number of each resource type label, it is specifically used to:

[0257] Determine the training weight of each initial loss function based on the number of labels for each resource type;

[0258] The target loss function is obtained by weighted summing the R initial loss functions according to the training weight of each initial loss function.

[0259] In a possible implementation manner, when the processor 1101 is used to determine the influencing factor of the object behavior feature and the influencing factor of the resource behavior feature according to the object attribute feature and the resource attribute feature, it is specifically used to:

[0260] Perform feature fusion on the object attribute features and the resource attribute features to obtain fused attribute features;

[0261] The influencing factors of the object behavior characteristics and the influencing factors of the resource behavior characteristics are generated according to the fused attribute characteristics.

[0262] In a possible implementation, the number of the object behavior features is N, where N is a positive integer greater than 1, and any one of the N object behavior features is represented as the i-th object behavior feature, where i is a positive integer less than or equal to N;

[0263] When the processor 1101 is used to determine the target object behavior feature according to the influencing factor of the object behavior feature and the object behavior feature, it is specifically used to:

[0264] Perform feature extraction on the i-th object behavior feature according to the N object behavior features to obtain an object extraction feature of the i-th object behavior feature;

[0265] The target object behavior feature corresponding to the ith object behavior feature is determined according to the object extraction feature of the ith object behavior feature and the influencing factor of the ith object behavior feature.

[0266] In a specific implementation, the above-described device, processor, memory, etc. can execute the implementation described in the above-mentioned method embodiment, and can also execute the implementation described in the embodiment of the present application, which will not be repeated here.

[0267] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0268] The embodiment of the present application also provides a computer program product, which includes program instructions, and when the program instructions are executed by a processor, some or all of the steps in the above method can be implemented. For example, the program instructions can be stored in a computer-readable storage medium. The processor of the computer device reads the program instructions from the computer-readable storage medium, and the processor executes the program instructions, so that the computer device performs some or all of the steps in the above method.

[0269] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from a website site, a computer, a server or a data center to another website site, a computer, a server or a data center by wired (e.g., coaxial cable, optical fiber, digital object line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or a data center that includes one or more available media integration. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)), etc.

[0270] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A data processing method, characterized in that: The method comprises: Obtain M sample information sets; one sample information set includes sample object attribute information and sample object behavior information of a sample object, and sample resource attribute information and sample resource behavior information of a sample recommended resource; each sample information set has a probability label and a resource type label, and the M sample information sets have R resource type labels, where M is a positive integer greater than 1, and R is a positive integer greater than 1 and less than or equal to M; Input each sample information set into the initial prediction model, and generate a sample probability corresponding to each sample information set based on the initial prediction model; any sample probability represents the probability of a sample object in the corresponding sample information set responding to a recommended behavior of a sample recommended resource in the sample information set; Determine a target loss function for the M sample information sets according to the number of each resource type label, train the initial prediction model according to the target loss function, the sample probability corresponding to each sample information set and the corresponding probability label, and obtain a target prediction model; Acquire a recommendation information set; the recommendation information set includes object attribute information and object behavior information of the target object, resource attribute information and resource behavior information of the resource to be recommended; Generate a recommended feature set according to the recommended information set; the recommended feature set includes object attribute features of the object attribute information, object behavior features of the object behavior information, resource attribute features of the resource attribute information, and resource behavior features of the resource behavior information; Determine the influencing factor of the object behavior feature and the influencing factor of the resource behavior feature according to the object attribute feature and the resource attribute feature; Determine a target object behavior feature according to the influencing factor of the object behavior feature and the object behavior feature, and determine a target resource behavior feature according to the influencing factor of the resource behavior feature and the resource behavior feature; Determine a target probability of the target object responding to the recommended behavior of the resource to be recommended according to the object attribute characteristics, the resource attribute characteristics, the target object behavior characteristics and the target resource behavior characteristics, and push the resource to be recommended to the object terminal of the target object according to the target probability; the target probability is generated by the target prediction model; The step of generating the sample probability corresponding to each sample information set based on the initial prediction model includes: Generate a sample feature set corresponding to the j-th sample information set based on the initial prediction model; the sample feature set includes sample object attribute features of the sample object attribute information, sample object behavior features of the sample object behavior information, sample resource attribute features of the sample resource attribute information, and sample resource behavior features of the sample resource behavior information in the j-th sample information set; the j-th sample information set is any one of the M sample information sets, and j is a positive integer less than or equal to M; Generating a sample impact factor of the sample object behavior feature and a sample impact factor of the sample resource behavior feature according to the sample object attribute feature and the sample resource attribute feature; Determine a target sample object behavior feature according to the sample influence factor of the sample object behavior feature and the sample object behavior feature, and determine a target sample resource behavior feature according to the sample influence factor of the sample resource behavior feature and the sample resource behavior feature; The sample probability corresponding to the j-th sample information set is determined according to the object attribute characteristics, the resource attribute characteristics, the target sample object behavior characteristics and the target sample resource behavior characteristics.

2. The method according to claim 1, characterized in that The initial prediction model includes R initial probability prediction networks, and one resource type label corresponds to one initial probability prediction network; The determining the sample probability corresponding to the j-th sample information set according to the object attribute feature, the resource attribute feature, the target sample object behavior feature, and the target sample resource behavior feature includes: In the jth initial probability prediction network, a sample probability corresponding to the jth sample information set is generated according to the object attribute characteristics, the resource attribute characteristics, the target sample object behavior characteristics and the target sample resource behavior characteristics; Among them, the j-th initial probability prediction network is the initial probability prediction network corresponding to the j-th resource type label in the R initial probability prediction networks, and the j-th resource type label is the resource type label that the j-th sample information set has in the R resource type labels.

3. The method according to claim 1, characterized in that The initial prediction model includes R initial probability prediction networks, and the R resource type labels are associated with R initial loss functions; and determining the target loss function for the M sample information sets according to the number of each resource type label includes: Obtaining a target probability prediction network from the R initial probability prediction networks; Determining the training weight of the target initial loss function associated with the resource type label corresponding to the target probability prediction network according to the number of resource type labels corresponding to the target probability prediction network; The target loss function is determined according to the training weight of the target initial loss function and the R initial loss functions.

4. The method according to claim 1, characterized in that The R resource type labels are associated with R initial loss functions; and the target loss function for the M sample information sets is determined according to the number of each resource type label, including: Determining the training weight of each initial loss function according to the number of each resource type label; The target loss function is obtained by weighted summing the R initial loss functions according to the training weight of each initial loss function.

5. The method according to claim 1, characterized in that The determining the influencing factor of the object behavior feature and the influencing factor of the resource behavior feature according to the object attribute feature and the resource attribute feature includes: Performing feature fusion on the object attribute feature and the resource attribute feature to obtain a fused attribute feature; An influencing factor of the object behavior feature and an influencing factor of the resource behavior feature are generated according to the fused attribute feature.

6. The method according to claim 1, characterized in that The number of the object behavior features is N, where N is a positive integer greater than 1, and any one of the N object behavior features is represented as the i-th object behavior feature, where i is a positive integer less than or equal to N; The determining the target object behavior feature according to the influencing factor of the object behavior feature and the object behavior feature comprises: Performing feature extraction on the i-th object behavior feature according to the N object behavior features to obtain an object extraction feature of the i-th object behavior feature; A target object behavior feature corresponding to the ith object behavior feature is determined according to the object extraction feature of the ith object behavior feature and the influencing factor of the ith object behavior feature.

7. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory is used to store a computer program, the computer program comprises program instructions, and the processor is configured to call the program instructions to execute the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor is caused to perform the method according to any one of claims 1 to 6.

Images