Data recommendation method, electronic device and storage medium

Abstract

The embodiment of the present application discloses a data recommendation 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 a first recommendation index of the target object for the resource to be recommended according to the recommendation feature set, determining a second recommendation index of the target object for the resource to be recommended under target scenario information according to the object attribute feature and the resource attribute feature, determining a target recommendation index of the target object for the resource to be recommended according to the first recommendation index and the second recommendation index, and pushing the resource to be recommended to the object terminal of the target object according to the target recommendation index. By adopting the embodiment of the present application, the efficiency and accuracy of recommendation prediction of the resource to be predicted can be improved.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and in particular to a data recommendation method, electronic device, and storage medium. Background Technology

[0002] The current purpose of recommendation tasks is primarily to predict resources that users are interested in within a target application and push them to the user's terminal, aiming to improve the effective recommendation rate for users. For example, predicting a streamer a user is interested in and recommending that streamer's live stream could increase the user's click-through rate. Existing prediction methods typically involve building prediction models to predict user-related attributes and resource-related attributes to achieve accurate recommendations. However, this approach relies on only these two attributes, resulting in a limited set of features learned by the model, leading to low recommendation prediction efficiency and accuracy. Therefore, improving the efficiency and accuracy of recommendation prediction for specific resources has become a pressing issue. Summary of the Invention

[0003] This application provides a data recommendation method, electronic device, and storage medium, which can improve the efficiency and accuracy of recommendation prediction for resources in multiple recommendation scenarios.

[0004] On the one hand, embodiments of this application provide a data recommendation method, which includes:

[0005] Obtain a set of recommendation information; this set of recommendation information includes the object attribute information of the target object, the resource attribute information of the resource to be recommended, and the recommendation scenario information of the resource to be recommended.

[0006] A set of recommendation features is generated based on the set of recommendation information; this set of recommendation features includes object attribute features of object attribute information, resource attribute features of resource attribute information, and recommendation scenario features of recommendation scenario information.

[0007] The first recommendation index of the target object for the resource to be recommended is determined based on the set of recommendation features; the first recommendation index represents the probability that the target object responds to the recommendation behavior for the resource to be recommended.

[0008] The second recommendation index of the target object for the resource to be recommended is determined based on the object attribute characteristics and resource attribute characteristics; the target scenario information is at least one of multiple scenario information associated with the resource to be recommended; the second recommendation index represents the probability that the target object responds to the recommendation behavior for the resource to be recommended.

[0009] The target recommendation index for the target object is determined based on the first recommendation index and the second recommendation index, and the resource to be recommended is pushed to the target object's terminal according to the target recommendation index; the target recommendation index represents the target probability of the target object responding to the recommendation behavior of the resource to be recommended.

[0010] On one hand, embodiments of this application provide a data recommendation device, which includes:

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

[0012] The processing module is used to generate a set of recommendation features based on the set of recommendation information; the set of recommendation features includes object attribute features of object attribute information, resource attribute features of resource attribute information, and recommendation scenario features of recommendation scenario information.

[0013] The determination module is used to determine a first recommendation index for a target object in relation to a resource to be recommended based on a set of recommendation features; the first recommendation index represents the probability that the target object will respond to the recommendation behavior for the resource to be recommended.

[0014] The determination module is used to determine a second recommendation index for a target object in relation to a resource to be recommended under target scenario information, based on object attribute features and resource attribute features; the target scenario information is at least one of multiple scenario information associated with the resource to be recommended, and the second recommendation index represents the probability that the target object responds to the recommendation behavior of the resource to be recommended;

[0015] The determination module is used to determine the target recommendation index of the target object for the resource to be recommended based on the first recommendation index and the second recommendation index; the target recommendation index represents the target probability of the target object responding to the recommendation behavior of the resource to be recommended.

[0016] The processing module is used to push the resources to be recommended to the target object's terminal according to the target recommendation index.

[0017] On one hand, embodiments of this application provide an electronic device including a processor and a memory, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute some or all of the steps in the above method.

[0018] On one hand, embodiments of this application provide a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform some or all of the steps in the above-described method.

[0019] Accordingly, according to one aspect of this application, a computer program product or computer program is provided, comprising program instructions stored in a computer-readable storage medium. A processor of a computer device reads the program instructions from the computer-readable storage medium and executes the program instructions, causing the computer device to perform the data recommendation method provided above.

[0020] In this embodiment, a set of recommendation information can be obtained and a set of recommendation features can be generated. A first recommendation index for the target object relative to the resource to be recommended is determined based on the set of recommendation features. This first recommendation index reflects the interaction of multiple recommendation features in the set of recommendation features. A second recommendation index for the target object under target scenario information is determined based on object attribute features and resource attribute features. This second recommendation index is obtained using learning features under target scenario information. A target recommendation index is determined based on the first and second recommendation indices, and the resource to be recommended is pushed to the target object's terminal according to the target recommendation index. The target recommendation index obtained in this way not only combines its own multiple recommendation features at a deeper level but also combines learning features under multiple recommendation scenarios. Therefore, it can improve the recommendation prediction efficiency and accuracy of the resource to be predicted, thereby improving the effect of subsequent recommendations based on the target recommendation index. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A schematic diagram of an application architecture provided for an embodiment of this application;

[0023] Figure 2 A flowchart illustrating a data recommendation method provided in an embodiment of this application;

[0024] Figure 3a This is a schematic diagram illustrating a recommended scenario information provided in an embodiment of this application.

[0025] Figure 3b This is a schematic diagram illustrating a recommended scenario information provided in an embodiment of this application.

[0026] Figure 3c This is a schematic diagram illustrating a recommended scenario information provided in an embodiment of this application.

[0027] Figure 4 A flowchart illustrating a data recommendation method provided in an embodiment of this application;

[0028] Figure 5 A schematic diagram illustrating a scenario for generating a set of recommendation features, provided as an embodiment of this application;

[0029] Figure 6a A schematic diagram illustrating a scenario for determining fusion attribute features, provided as an embodiment of this application;

[0030] Figure 6b A schematic diagram illustrating a scenario for determining fusion attribute features, provided as an embodiment of this application;

[0031] Figure 7 A schematic diagram illustrating a push scenario for recommended resources based on a target prediction model, provided as an embodiment of this application;

[0032] Figure 8 A flowchart illustrating a data recommendation method provided in an embodiment of this application;

[0033] Figure 9a A schematic diagram of a model training scenario provided in an embodiment of this application;

[0034] Figure 9b A schematic diagram of a prediction model provided in an embodiment of this application;

[0035] Figure 10 This is a schematic diagram of the structure of a data recommendation device provided in an embodiment of this application;

[0036] Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0037] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0038] The data recommendation method proposed in this application is implemented in an electronic device, which can be a server or a terminal. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, smart voice interaction device, smart home appliance, vehicle terminal, aircraft, etc., but is not limited to these.

[0039] In some embodiments, see Figure 1 , Figure 1 This is a schematic diagram of an application architecture provided for an embodiment of this application, through which the data recommendation method proposed in this application can be executed. Figure 1 As shown, the device may include an electronic device and a target terminal. The resource to be recommended belongs to a target application, and the target object may be a user (e.g., a user) targeted by the target application. The electronic device can obtain a set of recommendation information based on the target object and the resource to be recommended, and generate a set of recommendation features. This set of recommendation features includes a first recommendation feature of the target object, a second recommendation feature of the resource to be recommended, and a third recommendation feature of the recommendation scenario in which the resource to be recommended exists. Based on the aforementioned three recommendation features, a first recommendation index for the resource to be recommended is obtained. A second recommendation index is determined based on the first and second recommendation features and the recommendation scenario. A target recommendation index is then determined based on the first and second recommendation indices. Subsequently, a recommendation strategy for the resource to be recommended in the target terminal can be determined according to the target recommendation index, and the resource to be recommended is pushed to the target terminal based on this recommendation strategy. This is merely an example; the number of resources to be recommended can be one or more. Optionally, the above process can also be executed through a target prediction model. That is, a target prediction model can be deployed in the electronic device. In this target prediction model, a set of recommendation features is generated from the set of recommendation information, and a first recommendation index, a second recommendation index, and a target recommendation index are obtained based on this set of recommendation features.

[0040] Understandable, Figure 1 This is merely an illustrative representation of a possible application architecture for the technical solution of this application, and does not limit the specific architecture of the technical solution of this application. That is, the technical solution of this application may also provide other forms of application architecture.

[0041] Optionally, in some embodiments, the electronic device can execute the data recommendation method according to actual business needs to improve the prediction efficiency and accuracy of the resource to be predicted. That is, the electronic device can obtain the corresponding set of recommendation information based on the target object and the object to be recommended to generate a set of recommendation features. This set of recommendation features includes not only the relevant attribute features of the target object and the relevant attribute features of the resource to be predicted, but also the recommendation scene features corresponding to the recommendation scene information of the resource to be predicted. Based on the set of recommendation features and the recommendation scene of the object to be recommended, a first recommendation index and a second recommendation index are determined, thereby determining a target recommendation index. The object to be recommended can be pushed to the target object's terminal according to the target recommendation index, thereby achieving accurate push and improving the recommendation effect.

[0042] The recommended scenario information can be any scenario information associated with the resource to be recommended. For example, the scenario information could refer to the scenario where the resource to be recommended is recommended on the homepage of the target application to which it belongs; or it could refer to the scenario where the resource to be recommended is recommended on the display page of other resources, and so on.

[0043] Optionally, the technical solution of this application can be applied to any recommendation task. For example, it can be applied to live streaming recommendation tasks, where the object to be recommended can be the currently online streamer, and the target audience can be users of the live streaming application. The electronic device can determine the user's target recommendation index for the streamer based on the technical solution of this application, and push the streamer to the user's terminal based on the target recommendation index to increase the user's click-through rate and viewing time for the recommended streamer. As another example, it can also be applied to e-commerce recommendation tasks, where the object to be recommended can be a product, and the target audience can be users of the e-commerce application. The electronic device can determine the user's target recommendation index for the product based on the technical solution of this application, and push the product to the user's terminal based on the target recommendation index to increase the user's click-through rate and purchase rate for the recommended product.

[0044] Optionally, the data involved in this application, such as recommendation information sets and recommendation feature sets, can be stored in a database or in a blockchain, such as through a blockchain distributed system. This application does not impose any restrictions on this.

[0045] It is understood that the above scenarios are merely examples and do not constitute a limitation on the application scenarios of the technical solutions provided in the embodiments of this application. The technical solutions of this application can also be applied to other scenarios. For example, as those skilled in the art will know, with the evolution of system architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0046] Based on the above description, this application proposes a data recommendation method, which can be executed by the aforementioned electronic device. Please refer to... Figure 2 , Figure 2 This is a flowchart illustrating a data recommendation method provided in an embodiment of this application. Figure 2 As shown, the data recommendation method of this application embodiment may include the following flow:

[0047] S201. Obtain the set of recommendation information.

[0048] The recommendation information set may include the object attribute information of the target object, the resource attribute information of the resource to be recommended, and the recommendation scenario information of the resource to be recommended.

[0049] In some embodiments, the target object can be a user, and the object attribute information can include one or more user attributes that can be used to characterize the relevant characteristics of the user. The resource attribute information of the resource to be recommended can include one or more resource attributes that can be used to characterize the relevant characteristics of the resource to be recommended. The specific attribute types included in the object attribute information and resource attribute information can be set by relevant business personnel according to the recommendation task of the actual application, and are not limited here. For example, taking the application to live streaming recommendation task as an example, the object attribute information of the target object can include basic profile attributes (such as the user's age, gender, and education level), statistical attributes (such as the total duration of live streaming watched by the user in one day or three days, the total number of flowers sent by the streamer by the user in three days or one week, the total number of rewards given by the streamer by the user in three days or one week, etc.); the resource to be recommended can be the online streamer to be recommended, so the resource attribute information of the resource to be recommended can include basic profile attributes (such as the streamer's age, gender, and education level), real-time attributes (such as the total number of viewers or the highest number of viewers in the streamer's live room in one day or three days, the type of live room, the duration of the live room, etc.), etc.

[0050] It is understood that in the specific implementation of this application, user information such as user age and other related data are involved. 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 related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0051] In some embodiments, the resource to be recommended belongs to a target application in the target terminal. The resource can be recommended through multiple application interfaces within the target application. These multiple application interfaces can be considered as various recommendable scenario information associated with the resource in the target application. Furthermore, when recommending the resource, it can be displayed in different presentation formats across these multiple application interfaces. The multiple scenario information associated with the resource is related to the application to which the resource belongs.

[0052] Therefore, the recommendation scenario information of the resource to be recommended is any one of multiple scenario information associated with the resource to be recommended. The process of performing recommendation prediction for the resource to be recommended is the same under different recommendation scenario information, and the prediction resources used by the different recommendation scenario information in the recommendation prediction (such as relevant parameters involved in the process) can be partially shared and partially unique. This allows the technical solution of this application to realize recommendation prediction under multiple recommendation scenarios. Compared with modeling and predicting each recommendation scenario separately, this can combine the features of multiple recommendation scenarios to improve the effect of recommendation prediction. For example, the recommendation scenario information can be the application interface currently displayed in the target application, which is the scenario information in which the resource to be recommended is currently located and can be recommended. The electronic device can be the background device to which the target application belongs. When the target terminal detects that the target application displays a target application interface that can be recommended, it generates a recommendation instruction and sends it to the electronic device. When the electronic device receives the recommendation instruction, it can obtain the object attribute information of the target object and the resource attribute information of the resource to be recommended according to the instructions of the recommendation instruction, and determine the scenario information represented by the target application interface as the recommendation scenario information of the resource to be recommended.

[0053] The acquired resources to be recommended can be one or more. When there are multiple resources, the electronic device can construct a recommendation information set for each resource. Each recommendation information set includes the object attribute information of the target object, the resource attribute information of the resource to be recommended, and the current recommendation scenario information. The prediction process and principle for each recommendation information set are the same. The following explanation uses a single recommendation information set as an example.

[0054] For example, such as Figures 3a-3b As shown, Figures 3a-3b This is a schematic diagram of a recommendation scenario provided in an embodiment of this application; taking a live streaming recommendation task as an example, the resource to be recommended is a streamer in a live streaming application. The streamer can be recommended on the live streaming channel page of the live streaming application (i.e., scenario information 1). In this case, the streamer's live streaming room is represented by a live streaming tab and displayed in a thumbnail on the live streaming channel page, such as... Figure 3a As shown; when there are multiple streamers to recommend, the live channel page will display thumbnails of the live rooms of multiple streamers in sequence;

[0055] For example, streamer recommendations can also be made on the recommendation channel page of the live streaming application (i.e., scene information 2). This recommendation channel page directly displays the streamer's live stream details, and users can swipe up and down to enter the live streams of different recommended streamers, such as... Figure 3b As shown, each swipe represents a recommendation of a live streamer.

[0056] For example, streamer recommendations can also be made on the live streaming application's live streaming page (i.e., scene information 3). The target application, in response to a user's trigger action on streamer 1, displays the live stream details screen for streamer A. Live feed recommendations can be inserted into this details screen; that is, other streamers can be recommended in a designated area of ​​the details screen, such as... Figure 3c As shown, a thumbnail of streamer 2's live stream can be displayed in the lower left corner as a floating overlay. This live stream thumbnail can be displayed alongside... Figure 3a The thumbnails in the recommendations can be the same or different; in this case, the recommendation method can be to recommend a streamer once within a specified time, such as once every 10 seconds, with a maximum of 10 streamers recommended.

[0057] It is understood that the technical solution of this application can be applied to any recommendation task. Depending on the recommendation task, the target object, the resource to be recommended, and the recommendation scenario information can all be different. There are no restrictions on the recommendation task here. For ease of explanation, the following will use live recommendation task as an example.

[0058] S202. Generate a set of recommendation features based on the set of recommendation information.

[0059] The recommendation feature set can include object attribute features of object attribute information, resource attribute features of resource attribute information, and recommendation scenario features of recommendation scenario information.

[0060] In one possible embodiment, the generation process and principle of each recommendation feature in the recommendation feature set are the same; here, object attribute features are taken as an example. Specifically, the electronic device generates object attribute features based on object attribute information by performing one-hot encoding on the object attribute information to obtain an encoded feature vector, and then using this encoded feature vector as the object attribute feature. For example, if the object attribute information includes an age attribute, and this age attribute is divided into multiple attribute categories [<18, 19-30, 31-40, 41-50, 51-60, >60], and the target object's age is 24, then the target attribute category is [19-30]. Therefore, the encoded feature vector corresponding to the age attribute obtained by one-hot encoding can be represented as [0,1,0,0,0,0].

[0061] Therefore, corresponding encoded feature vectors can be generated from one or more attributes contained in the object attribute information, and the object attribute feature can be composed of these one or more encoded feature vectors. During hot-coded identification, the method for dividing attributes can be set by relevant business personnel based on experience. The specific methods for generating resource attribute features and recommended scenario features for electronic devices are the same as those for generating object attribute features, and will not be elaborated here.

[0062] In one possible embodiment, the electronic device can generate object attribute features based on object attribute information by performing one-hot encoding on the object attribute information to obtain an encoded feature vector, and then obtaining the embedded feature vector corresponding to the encoded feature vector, using the embedded feature vector as the object attribute feature. Specifically, the electronic device can obtain the embedded feature vector corresponding to the encoded feature vector by acquiring an embedded feature matrix constructed for the object attribute information, and determining the corresponding embedded feature vector from the embedded feature matrix according to the encoded feature vector and a preset method.

[0063] In some embodiments, the electronic device determines the corresponding embedded feature vector from the embedded feature matrix according to the encoded feature vector and in a preset manner. Specifically, this may involve multiplying the encoded feature vector with the embedded feature matrix and using the product as the corresponding embedded feature vector. Since the encoded feature vector consists of 0s and 1s, the resulting embedded feature vector indicates that the column value of the element with a value of 1 in the encoded feature vector corresponds to the target row vector of the embedded feature matrix. Therefore, corresponding embedded feature vectors can be generated from one or more attributes contained in the object attribute information, and the object attribute features can be composed of one or more embedded feature vectors.

[0064] For example, if an object's attribute information includes an age attribute, and the encoded feature vector obtained by hot-coding this age attribute is [0,1,0,0,0,0], then the column containing the element with a value of 1 has a value of 2. Therefore, the vector in the second row of the embedded feature matrix is ​​taken as the embedded feature vector corresponding to this age attribute. The embedded feature matrix constructed when determining the embedded feature vector can be set by relevant business personnel based on experience. The embedded feature matrices corresponding to different attribute information can be the same or different, and there is no limit to the size of the embedded feature matrix. The specific methods for generating resource attribute features and recommended scenario features for electronic devices are the same as those for generating object attribute features, and will not be elaborated here.

[0065] Optionally, the object attribute features, resource attribute features, and recommendation scenario features in the above-mentioned recommendation information set can all simultaneously include their respective encoded feature vectors and embedded feature vectors.

[0066] S203. Determine the first recommendation index of the target object for the resource to be recommended based on the set of recommendation features.

[0067] Among them, the first recommendation index can characterize the probability that the target object will respond to the recommended resource.

[0068] In some embodiments, when a target object is interested in a recommended resource, it can respond to the recommendation behavior for that resource. This recommendation behavior can be understood as generating an effective recommendation for the target object using that resource. The first recommendation index can also represent the target object's level of interest in the recommended resource. Therefore, the larger the value indicated by the first recommendation index, the greater the probability that the target object will respond to the recommendation behavior, the greater the probability that the recommended resource will generate an effective recommendation, and the greater the target object's level of interest in the recommended resource. The specific type of recommendation behavior is related to the recommendation scenario information of the resource to be recommended. Different recommendation scenario information may result in different recommendation behaviors; therefore, the type of recommendation behavior is not limited here.

[0069] For example, taking the application of live streaming recommendation tasks as an example, if the recommendation scenario information is... Figure 3a When the indicated scenario information is provided, the recommendation behavior can be clicking on a streamer's live broadcast room on the recommended channel page. In this case, the first recommendation index can represent the predicted click-through rate of the target audience for the recommended streamer; if the recommendation scenario information is... Figure 3b When the indicated scenario information is provided, the recommendation behavior can be watching a streamer's live broadcast room on the recommended channel page. For example, watching streamer A's live broadcast room for 3 seconds indicates that streamer A has generated a valid recommendation. In this case, the first recommendation index can represent the target audience's predicted viewing rate of the recommended streamer. If the recommendation scenario information is... Figure 3c When the indicated scenario information is provided, the recommended behavior can be the clicking behavior of other streamers recommended on the live streaming page. In this case, the first recommendation index can represent the predicted click-through rate of the target object for the recommended streamer.

[0070] In some embodiments, the electronic device may obtain a first recommendation index based on a combination of independent recommendation features and fused recommendation features from a set of recommendation features. Therefore, the first recommendation index includes not only directly displayed feature information obtained from the independent recommendation features, but also implicitly displayed feature information based on the fused recommendation features. The specific method for obtaining the first recommendation index based on the combination of independent recommendation features and fused recommendation features from the set of recommendation features can be found in the following description of the embodiments.

[0071] In some embodiments, the electronic device determines a first recommendation index for a target object relative to a resource to be recommended based on a set of recommended features, which can be obtained through a target prediction model. Specifically, this can involve acquiring a target prediction model and predicting the first recommendation index within the model based on object attribute features, resource attribute features, and recommendation scenario features; where the object attribute features, resource attribute features, and recommendation scenario features each include their respective encoded feature vectors and embedded feature vectors. When there are multiple sets of recommended features, the model parameters used in the target prediction model for predicting these multiple sets can be the same; that is, when the first recommendation index is obtained, each set of recommended features shares a portion of the model network of the target prediction model to share model information. This target prediction model is trained based on the process described in steps S801-S804 of the following embodiments.

[0072] S204. Determine the second recommendation index of the target object for the resource to be recommended under the target scenario information based on the object attribute characteristics and resource attribute characteristics.

[0073] Among them, the target scene information is at least one of the multiple scene information associated with the resource to be recommended, and the second recommendation index can characterize the probability that the target object responds to the recommended resource in a way that is associated with the target scene information.

[0074] In some embodiments, the second recommendation index can represent the degree of interest of the target object in the recommended resource under the target scenario information. Therefore, the larger the value indicated by the second recommendation index, the greater the probability that the target object will respond to the recommendation behavior under the target scenario information, the greater the probability that the recommended resource will generate an effective recommendation under the target scenario information, and the greater the degree of interest of the target object in the recommended resource under the target scenario information.

[0075] In one possible implementation, the recommended scenario information is the scenario information of the resource to be recommended in the target application. The multiple scenario information associated with the resource to be recommended includes the recommended scenario information, and the target scenario information can be determined based on the recommended scenario information.

[0076] In some embodiments, the target scene information can be recommended scene information, or it can be related scene information of recommended scene information and multiple scene information. The related scene information of recommended scene information can be all scene information other than recommended scene information in multiple scene information, or it can be a specific scene information in multiple scene information, which can be set by relevant business personnel based on experience values.

[0077] For example, let the recommended scenario information be as follows: Figure 3bThe indicated scene information 2, at this time the recommended scene information is the application interface playing the live broadcast screen, and such as Figure 3c The application interface indicated by scene information 3 also plays live broadcast footage, so scene information 3 can be set as the associated scene information of scene information 2.

[0078] In some embodiments, the second recommendation index can be obtained through the target prediction model described above, which may include multiple scene prediction networks. Each scene prediction network corresponds to one scene information, meaning each scene information has a unique scene prediction network.

[0079] When the target scene information is the recommended scene information, the electronic device determines the second recommendation index of the target object for the resource to be recommended under the target scene information based on the object attribute features and resource attribute features. Specifically, this can be achieved by predicting the object attribute features and resource attribute features based on the scene prediction network corresponding to the recommended scene information to obtain the second recommendation index. Here, the object attribute features and resource attribute features predicted are the corresponding embedded feature vectors.

[0080] When the target scene information consists of recommended scene information and associated scene information, the electronic device determines a second recommendation index for the target object relative to the resource to be recommended under the target scene information based on object attribute features and resource attribute features. Specifically, this can be achieved by predicting the aforementioned object attribute features and resource attribute features using the scene prediction networks corresponding to the recommended scene information and the associated scene information, respectively, to obtain the second recommendation index. The specific method for obtaining the second recommendation index using the scene prediction networks corresponding to the recommended scene information and the associated scene information can be found in the relevant description of the following embodiments. Here, the predicted object attribute features and resource attribute features are all corresponding embedded feature vectors.

[0081] It is understood that the aforementioned scene prediction networks are trained using object attribute features and resource attribute features under their respective scene information, thus containing unique learned features under their respective scene information. During prediction, object attribute features and resource attribute features can be assigned to corresponding dedicated scene prediction networks based on the target scene information to obtain a second recommendation index. This second recommendation index contains the feature information of object attribute features and resource attribute features under the premise of the learned features of the target scene information. When the target prediction model predicts different recommendation feature sets, the recommendation scene information in the recommendation feature sets is different, and the scene prediction networks used are also different. That is, the model parameters used during prediction can be different and different from the model parameters used for the first recommendation index. Therefore, although both the first and second recommendation indices are based on recommendation feature sets, they are obtained from different prediction perspectives using different model parameters. Based on the above description, the target prediction model can at least include a common prediction network for predicting the first recommendation index and multiple scene prediction networks for predicting the second recommendation index.

[0082] S205. Determine the target recommendation index for the resource to be recommended for the target object based on the first recommendation index and the second recommendation index, and push the resource to be recommended to the target object's terminal according to the target recommendation index.

[0083] The target recommendation index can be obtained by combining the first and second recommendation indices, which can improve the predictive accuracy of the target recommendation index. This target recommendation index represents the target probability of a target object responding to the recommended resource.

[0084] In some embodiments, the electronic device determines the target recommendation index for the target object relative to the resource to be recommended based on the first recommendation index and the second recommendation index. Specifically, this can be achieved by weighted summation of the first recommendation index and the second recommendation index to obtain the target recommendation index. The weighting coefficients can be set by relevant business personnel.

[0085] Optionally, the target recommendation index can also be determined based on the first and second recommendation indices through a target prediction model. The target recommendation index can be output from the predicted first and second recommendation indices in the target prediction model. For example, the aforementioned weighting coefficients can also be obtained as model parameters in the target prediction model based on training.

[0086] In some embodiments, the target recommendation index can represent the target object's level of interest in the recommended resources under the current recommendation scenario information. Therefore, a recommendation strategy for the resources to be recommended can be determined based on the target recommendation index, and the resources to be recommended can be pushed to the target terminal based on the recommendation strategy. This recommendation strategy can be set by relevant business personnel based on experience. For example, when there are multiple resources to be recommended, the recommendation strategy can be to push them to the target terminal in descending order of the target recommendation index; or, the recommendation strategy can also be to push the resource to be recommended when the target recommendation index is greater than or equal to a preset index, and not push it when it is less than the preset index, and so on.

[0087] In this embodiment, a set of recommendation information can be obtained, a set of recommendation features can be generated based on the set of recommendation information, and a first recommendation index for the target object relative to the resource to be recommended can be determined based on the set of recommendation features. This first recommendation index reflects the interaction of multiple recommendation features in the set of recommendation features. A second recommendation index for the target object relative to the resource to be recommended can be determined based on object attribute features and resource attribute features. This second recommendation index is obtained using learning features under the target scene information, which can be obtained by combining learning features from all scene information. A target recommendation index for the target object relative to the resource to be recommended can be determined based on the first and second recommendation indices, and the resource to be recommended can be pushed to the target object's terminal according to the target recommendation index. The target recommendation index obtained in this way can not only combine its own multiple recommendation features at a deeper level, but also combine learning features under multiple recommendation scenarios, which can improve the recommendation prediction efficiency and accuracy of the resource to be predicted, and is conducive to improving the effect of subsequent recommendation based on the target recommendation index.

[0088] Please see Figure 4 , Figure 4 This is a flowchart illustrating a data recommendation method provided in an embodiment of this application. This method can be executed by the aforementioned electronic device. Figure 4 As shown, the data recommendation method in this embodiment may include the following steps:

[0089] S401. Obtain the recommended information set. The specific implementation of step S401 can be found in the relevant descriptions of the above embodiments, and will not be repeated here.

[0090] S402. Generate a set of recommendation features based on the set of recommendation information.

[0091] The recommendation feature set includes object attribute features of object attribute information, resource attribute features of resource attribute information, and recommendation scenario features of recommendation scenario information.

[0092] In one possible implementation, the generation process and principle of each recommendation feature in the recommendation feature set are the same, and each recommendation feature can include the encoded feature vector and the embedded feature vector obtained from each recommendation information in the recommendation information set. The specific method for generating the recommendation feature set can be found in the relevant description of the above embodiments.

[0093] For example, such as Figure 5 As shown, Figure 5 This is a schematic diagram of a scenario for generating a set of recommendation features according to an embodiment of this application. The object attribute information includes object attribute 1, and the encoded feature vector corresponding to object attribute 1 is [0,1,0,0,0,0] (denoted as U1). Therefore, the row vector of the second row can be obtained from the corresponding embedding matrix 1 as the embedded feature vector corresponding to object attribute 1 (denoted as V1). Based on the object attribute information (including object attribute 1, object attribute 2, ..., object attribute n), the corresponding object attribute features are obtained as encoded feature vectors [U1, U2, ..., Un] and embedded feature vectors [V1, V2, ..., Vn]. Based on the resource attribute information (including resource attribute n+1, resource attribute n+2, ..., resource attribute N), the corresponding resource attribute features are obtained as encoded feature vectors [Un+1, Un+2, ..., UN] and embedded feature vectors [Vn+1, Vn+2, ..., VN]. Based on the recommendation scenario information (denoted as m), the corresponding recommendation scenario features are obtained as encoded feature vector [Um] and embedded feature vector [Vm].

[0094] Optionally, the above process can also be obtained through a target prediction model, that is, the target prediction model may include a feature generation layer, which generates a set of recommended features based on the set of recommended information. This embedded feature matrix can be used as model parameters in the target prediction model. This target prediction model is trained based on the process described in steps S801-S804 of the following embodiments.

[0095] S403. Determine the fusion attribute features and recommendation impact value based on the recommendation feature set, and determine the first recommendation index based on the fusion attribute features and recommendation impact value.

[0096] In one possible implementation, the electronic device can determine the recommendation influence value based on multiple independent recommendation features in the recommendation feature set, and perform feature fusion based on the multiple independent recommendation features to obtain fused attribute features. The fused attribute features can deepen the feature interaction between multiple recommendation features to obtain more feature information.

[0097] In one possible implementation, the recommendation impact value can characterize the degree of influence of various recommendation features in the recommendation information set on the target object's response to the recommendation behavior. The electronic device determines the recommendation impact value based on the recommendation feature set using the aforementioned target prediction model. This target prediction model may further include a common prediction network, which may include an influence prediction network. The electronic device can obtain all encoded feature vectors from the recommendation feature set, and the influence prediction network predicts the recommendation impact value based on all encoded feature vectors. This encoded feature vector consists of 0s and 1s, with only the element corresponding to the target attribute category based on the attribute being classified having a value of 1. Therefore, the prediction of the recommendation impact value can be performed using this encoded feature vector, indicating that the prediction is based solely on the classified target attribute category.

[0098] Optionally, the recommendation impact value can be positive or negative. When the recommendation impact value is positive and larger, the positive impact of the multiple recommendation features is greater; conversely, when the recommendation impact value is negative and larger, the negative impact of the multiple recommendation features is greater. For example, if the recommendation impact value of the target object on the recommended resource A is determined to be 1 based on recommendation information set 1, and the recommendation impact value of the target object on the recommended resource B is determined to be 2 based on recommendation information set 2, then the target object is more likely to respond to the recommendation behavior of the recommended resource B under the influence of recommendation information set 2.

[0099] In one possible implementation, the fused attribute feature represents the generation of a new fused recommendation feature from multiple recommendation features in the recommendation feature set; this can obtain the most differentiated information from the multiple original recommendation features involved in the fusion process and can eliminate redundant information caused by the correlation between different recommendation features. The electronic device determining the fused attribute feature based on the recommendation feature set can be achieved by the aforementioned target prediction model. This target prediction model may include a common prediction network, which may include a feature fusion network.

[0100] In some embodiments, the electronic device utilizes a feature fusion network to perform feature fusion based on a recommended feature set to obtain fused attribute features. Specifically, this may involve obtaining all embedded feature vectors in the recommended feature set and performing feature fusion on all embedded feature vectors.

[0101] In some embodiments, the number T of all embedded feature vectors (denoted as V) includes a first embedded feature vector (denoted as Vi) and a second embedded feature vector (denoted as Vj). Therefore, feature fusion of all embedded feature vectors can be specifically performed according to the following fusion formula:

[0102] P ij =V i *W*V j T *wij Formula 1.1;

[0103] In the above, W is the common weight matrix; w ij These are unique weight parameters for the first and second embedded feature vectors; V j T Let i be the transpose of the second embedded feature vector; i and j are both positive integers from 1 to T, i can be equal to j or not equal to j; when i equals j, P ij Default parameters can be set for relevant business personnel.

[0104] Here, let the number of elements in the first embedded feature vector be n, and the number of elements in the second embedded feature vector be m. Then, the size of the common weight matrix is ​​n x m. Since the size of the embedded feature matrix corresponding to each attribute can be different, the elements of the resulting embedded feature vectors can also be different. Multiple common weight matrices can be used during feature fusion, each with a different size determined based on the two embedded feature matrices being fused. When the first and second embedded feature vectors are fused, a matching common weight matrix can be selected from multiple common weight matrices based on the number of elements in the first and second embedded feature vectors, respectively.

[0105] Alternatively, there can be only one target weight matrix. The size of the target weight matrix is ​​determined based on the number of elements in the target embedded feature vector. The target embedded feature vector is the feature vector with the most elements among all embedded feature vectors. Let the number of elements be Nmax. Then the size of the target weight matrix can be an Nmax row and Nmax column matrix. When the first embedded feature vector and the second embedded feature vector are fused, nxm weight values ​​are extracted from the target weight matrix starting from the top left corner of the matrix to form the required common weight matrix.

[0106] Based on the above fusion formula, a fusion element can be obtained from every two embedded feature vectors. Multiple fusion elements can be obtained from all the embedded feature vectors, and the feature matrix formed by these multiple fusion elements in sequence is the fusion feature matrix. This fusion formula references the idea of ​​WKFM (a weighted FM (Factorization Machine)). Optionally, the above common weight matrix and unique weight parameters can be set by relevant business personnel based on experience, or they can all be network parameters in the feature fusion network, obtained by training the target prediction model.

[0107] For example, such as Figures 6a-6b As shown, Figures 6a-6bThis application provides a schematic diagram of a scenario for determining fused attribute features; wherein, the recommended feature set includes embedded feature vector 1 (V1, with 3 elements), embedded feature vector 2 (V2, with 4 elements), and embedded feature vector 3 (V3, with 5 elements); therefore:

[0108] like Figure 6a Suppose there exist multiple common weight matrices, the size of which is determined based on the number of elements in each embedded feature vector. These matrices can be 3x4, 3x5, ..., etc. During feature fusion: when embedding feature vector 1 and embedding feature vector 2 are fused, a common weight matrix of size 3x4 is selected from the multiple common weight matrices based on the number of elements in each feature vector; when embedding feature vector 1 and embedding feature vector 3 are fused, a common weight matrix of size 3x5 is selected from the multiple common weight matrices based on the number of elements in each feature vector.

[0109] For example, such as Figure 6b When a target weight matrix exists, it can be a 5x5 matrix. During feature fusion: when embedding feature vector 1 and embedding feature vector 2 are fused, 3x4 weight values ​​are extracted sequentially from the target weight matrix based on the number of elements in embedding feature vector 1 and embedding feature vector 2 to form a common weight matrix 1; when embedding feature vector 1 and embedding feature vector 3 are fused, 3x5 weight values ​​are extracted sequentially from the target weight matrix based on the number of elements in embedding feature vector 1 and embedding feature vector 3 to form a common weight matrix 2.

[0110] Therefore, embedded feature vector 1 is sequentially coupled with the common weight matrix 1, the transpose of embedded feature vector 2, and the unique weight parameters (w) corresponding to embedded feature vector 1 and embedded feature vector 2 according to the above fusion formula. 12 Multiplying them together yields the fused element (let's call it P). 12 ); Embedded feature vector 1 is sequentially coupled with the common weight matrix 2, the transpose of embedded feature vector 3, and the unique weight parameters (w) corresponding to embedded feature vector 1 and embedded feature vector 3 according to the above fusion formula. 13 Multiplying them together yields the fused element (let's call it P). 13 Based on V1, V2 and V3, nine fusion elements can be obtained, which are then arranged in sequence to form a 3*3 fusion feature matrix; this fusion feature matrix is ​​the fusion attribute feature.

[0111] In some embodiments, the electronic device may determine a first recommendation index based on fused attribute features and recommendation influence values ​​according to a preset determination method. This preset determination method can be set by relevant business personnel when constructing the target prediction model.

[0112] For example, the preset determination method could be to obtain the first recommendation index by weighted summation of the sum of each element in the fused attribute features and the recommendation influence value; or, the preset determination method could be to input the fused attribute features into a fused feature prediction network to obtain a predicted value, and then determine the first recommendation index by the sum of the predicted value and the recommendation influence value, in which case the target prediction model could also include the fused feature prediction network; or, the preset determination method could be to input the fused attribute features and the recommendation influence value together into an index prediction network to obtain the first recommendation index, in which case the target prediction model could also include the index prediction network; and so on. No limitation is made to the preset determination method here.

[0113] S404. Determine the second recommendation index of the target object for the resource to be recommended under the target scenario information based on the object attribute characteristics and resource attribute characteristics.

[0114] The target scenario information is at least one of multiple scenario information associated with the resource to be recommended, and these multiple scenario information may include the recommendation scenario information of the resource to be recommended. The second recommendation index represents the probability that the target object will respond to the recommendation behavior of the resource to be recommended.

[0115] In one possible implementation, the second recommendation index for target scene information determined by the electronic device based on object attribute characteristics and resource attribute characteristics can be obtained by the aforementioned target prediction model. This target prediction model may include a scene prediction network for multiple scene information, with one scene prediction network corresponding to one scene information.

[0116] In some embodiments, the target scene information may be recommended scene information; the electronic device may acquire multiple scene prediction networks in the target prediction model, and predict the embedded feature vectors of object attribute features and resource attribute features based on the scene prediction networks corresponding to the recommended scene information in the multiple scene prediction networks, thereby obtaining a second recommendation index.

[0117] In some embodiments, the target scene information may include recommended scene information and associated scene information of the recommended scene information among multiple scene information. There may be one or more associated scene information, which can be set by relevant business personnel based on experience values. The electronic device can acquire multiple scene prediction networks in the target prediction model, and predict object attribute features and resource attribute features based on the scene prediction networks corresponding to the recommended scene information and the associated scene information, respectively. This yields a first initial recommendation index for the target object regarding the resource to be recommended in the recommended scene information and a second initial recommendation index for the target object regarding the resource to be recommended in the associated scene information. When there are multiple associated scene information, there are also multiple corresponding second initial recommendation indices. The second recommendation index is obtained based on the first and second initial recommendation indices.

[0118] In some embodiments, the electronic device obtains the second recommendation index based on the first and second initial recommendation indices by performing a weighted summation of the first and second initial recommendation indices. The aforementioned weighting coefficients can be set by relevant business personnel based on experience.

[0119] S405. Determine the target recommendation index for the target object relative to the resource to be recommended based on the first recommendation index and the second recommendation index. The specific implementation of step S405 can be found in the relevant description of the above embodiments, and will not be repeated here.

[0120] S406. Push the resources to be recommended to the target object's terminal according to the target recommendation index.

[0121] In some embodiments, electronic devices can determine the recommended resources that the target object is most likely to be interested in based on a target recommendation index, thereby improving the effective recommendation rate. Therefore, pushing recommended resources to the target object's terminal according to the target recommendation index can be done by pushing recommended resources with a target recommendation index greater than a preset index; or it can be done by pushing recommended resources in descending order of the target recommendation index; and so on.

[0122] For example, such as Figure 7 This application provides a schematic diagram of a push scenario for recommended resources based on a target prediction model, taking its application to live streaming recommendation tasks as an example; wherein:

[0123] When the target application on the user terminal displays the application interface indicated by any of the multiple scene information, the electronic device obtains multiple currently online broadcasters (i1, i2, i3, ..., iN), and forms prediction pairs ([u,i1,s], [u,i2,s], [u,i3,s], ..., [u,iN,s]) with the target user (u) and each broadcaster and the currently present recommended scene information (s). Based on these prediction pairs, the device extracts the user attribute information of the target user and the broadcaster attribute information of each broadcaster from the attribute storage platform (such as a database), and combines the user attribute information with the broadcaster attribute information to form prediction pairs. Information pairs are generated and each information pair is sequentially input into the target prediction model. In the target prediction model, feature pairs corresponding to each information pair are generated (including user attribute features of user attribute information, anchor attribute features of corresponding anchor attribute information, and recommendation scene features of recommendation scene information). The first recommendation index and the second recommendation index corresponding to each feature pair are obtained respectively. The target recommendation index of each feature pair is determined based on the corresponding first recommendation index and the second recommendation index. Each anchor can be sorted according to the target recommendation index of each feature pair to obtain the sorted anchors. The sorted anchors or a target number of sorted anchors are sequentially pushed to the user terminal of the target user.

[0124] Recommendations based on the above-ranked streamers can specifically include, for example: Figure 3a As shown, the system sequentially retrieves the thumbnails of the live streams of the streamers (e.g., streamer 1, streamer 3, streamer 2, ...) and displays these thumbnails sequentially on the live stream channel page of the target application; for example... Figure 3b As shown, the program sequentially retrieves the live stream details of the sorted streamers (e.g., streamer 1, streamer 3, streamer 2, ...), and displays the live stream details of streamer 1 on the target application's recommendation channel page. Users can enter the live stream of streamer 1 by tapping "Enter Live Stream". When the target user's swipe operation is detected, the program sequentially displays the live stream details of streamer 3 on the recommendation channel page, indicating that the recommendation behavior has been executed.

[0125] In this embodiment, a recommendation information set can be obtained, a recommendation feature set can be generated based on the recommendation information set, fusion attribute features and recommendation influence values ​​can be determined based on the recommendation feature set, and a first recommendation index can be determined based on the fusion attribute features and recommendation influence values. This first recommendation index, based on the fusion attribute features, reflects the interaction of multiple recommendation features in the recommendation feature set. Furthermore, the first recommendation index, also based on the recommendation influence values, reflects the degree of influence of multiple recommendation features on the response recommendation behavior. A second recommendation index is determined based on object attribute features and resource attribute features for the target object in the target scene information. This second recommendation index is obtained using learned features in the target scene information. The learning features under scene information can be obtained by comprehensively learning features from all scene information; the target recommendation index for the target object is determined based on the first recommendation index and the second recommendation index, and the resource to be recommended is pushed to the target object's terminal according to the target recommendation index; the target recommendation index obtained in this way can not only combine its own multiple recommendation features at a deeper level, but also combine the learning features under multiple recommendation scenarios, which can improve the recommendation prediction efficiency and accuracy of the resource to be predicted, which is conducive to improving the effect of subsequent recommendation based on the target recommendation index. In addition, the recommendation scene information can be any scene information among the multiple associated scene information, so recommendation prediction under multiple recommendation scenarios can be realized.

[0126] Please see Figure 8 , Figure 8 This is a flowchart illustrating a data recommendation method provided in an embodiment of this application. This method can be executed by the aforementioned electronic device. Figure 8 As shown, the data recommendation method in this embodiment may include the following steps:

[0127] S801, Obtain a set of multiple sample information.

[0128] A sample information set includes the sample object attribute information of a sample object, the sample resource attribute information of a sample recommended resource, and the sample scene information of the sample recommended resource. The sample object, sample recommended resource, and sample scene information in multiple sample information sets can all be different.

[0129] In some embodiments, the aforementioned target recommendation index can be generated by a target prediction model. Therefore, the electronic device can use these multiple sample information sets as a batch of sample data to train the initial prediction model in batches, and perform multiple rounds of batch training until convergence to obtain the final target prediction model. The process and principle of each round of batch training are the same; here, the k-th round of training is used as an example. The sample objects, sample recommendation resources, and sample scene information in the multiple sample information sets of the k-th round can all be different. The sample data contained in this sample information set can be the same data used in the application process described in the above embodiments.

[0130] S802. Input each sample information set into the initial prediction model, and generate a sample feature set corresponding to each sample information set based on the initial prediction model.

[0131] Among them, a sample feature set includes sample object attribute features of sample object attribute information, sample resource attribute features of sample resource attribute information, and sample recommendation scenario features of sample scenario information of the sample resource attribute information.

[0132] In some embodiments, the initial prediction model may include a feature generation layer to be trained. The model parameters in the feature generation layer may include an embedding feature matrix corresponding to each attribute, which can be obtained by model training. The electronic device can sequentially input each set of sample information into the initial prediction model, generate an encoded feature vector and an embedded feature vector corresponding to each set of sample information based on the feature generation layer in the initial prediction model, and use these corresponding encoded feature vectors and embedded feature vectors as the sample feature set corresponding to each set of sample information. The specific method by which the encoded feature vector and embedded feature vector corresponding to each set of sample information are generated in the feature generation layer based on the embedding feature matrix corresponding to each attribute can be the same as described in the above embodiments.

[0133] S803. Determine the target sample recommendation index for the sample recommendation resources corresponding to each sample feature set based on each sample feature set.

[0134] In some embodiments, the process and principle of the electronic device determining the target sample recommendation index corresponding to each sample information set based on each sample feature set are the same. Here, we take any sample feature set in each sample feature set as an example for explanation. Let any sample feature set be the target feature set, the target sample object be the sample object corresponding to the target feature set, the target sample recommendation resource be the sample recommendation resource corresponding to the target sample object, and the target sample scene information be the sample scene information corresponding to the target sample recommendation resource.

[0135] Therefore, the electronic device determines the target sample recommendation index for each sample object corresponding to each sample feature set based on the sample recommendation resources corresponding to that sample object. Specifically, in the initial prediction model, a first sample recommendation index for the target sample object to the target sample recommendation resources is generated based on the target feature set, and a second sample recommendation index for the target sample object to the target sample recommendation resources is determined based on the sample object attribute features and sample resource attribute features in the target feature set; the target sample recommendation index for the target sample object to the target sample recommendation resources is determined based on the first and second sample recommendation indices.

[0136] Among them, the first sample recommendation index can characterize the initial sample probability of the target sample object responding to the recommendation behavior of the target sample recommendation resource; the second sample recommendation index can characterize the initial sample probability of the target sample object responding to the recommendation behavior of the target sample recommendation resource under the target sample scenario information; and the target sample recommendation index can characterize the target sample probability of the target sample object responding to the recommendation behavior of the target sample recommendation resource.

[0137] In one possible implementation, the initial prediction model may include an initial common prediction network and multiple initial scene prediction networks, with each initial scene prediction network corresponding to one sample scene information. Specifically, the electronic device generates a first sample recommendation index for the target sample object relative to the target sample recommendation resource based on the target feature set in the initial prediction model. Specifically, this can be achieved by generating the first sample recommendation index based on the target feature set in the initial common prediction network. The electronic device determines a second sample recommendation index for the target sample object relative to the target sample recommendation resource based on the sample object attribute features and sample resource attribute features in the target feature set. Specifically, this can be achieved by determining the second sample recommendation index in the target scene prediction network based on the embedded feature vectors of the sample object attribute features and the embedded feature vectors of the sample resource attribute features in the target feature set. This target scene prediction network is the initial scene prediction network corresponding to the target sample scene information among the multiple initial scene prediction networks.

[0138] In some embodiments, the aforementioned initial public prediction network may include an initial feature fusion network and an initial impact prediction network. The electronic device generates a first sample recommendation index based on the target feature set in the initial public prediction network by performing feature fusion on all embedded feature vectors in the target feature set to obtain sample fusion attribute features of the target information set; and in the initial impact prediction network, performing feature fusion on all encoded feature vectors in the target feature set to obtain sample recommendation impact values ​​of the target information set, summing all element values ​​in the sample fusion attribute features to obtain a sample fusion value, and using the sum of the sample fusion value and the sample recommendation impact value as the first sample recommendation index.

[0139] S804. Adjust the model parameters of the initial prediction model according to the recommendation index of each target sample to obtain the target prediction model.

[0140] In some embodiments, the electronic device obtains the target prediction model by correcting the model parameters of the initial prediction model based on the recommendation index of each target sample. Specifically, this can be achieved by correcting the network parameters of the initial common prediction network based on the recommendation index of each target sample to obtain a common prediction network; and by correcting the network parameters of multiple initial scene prediction networks based on the recommendation index of each target sample and the sample scene information of each sample recommendation resource to obtain multiple scene prediction networks; and by determining the target prediction model based on the common prediction network and the multiple scene prediction networks. The final target prediction model can be obtained by correcting the model parameters multiple times. If the model parameters of the initial prediction model also include the embedded feature matrix in the feature generation layer and the weighting coefficients used for weighted summation, then the aforementioned model parameters can be corrected using the recommendation index of each target sample, and the final target prediction model can be obtained through the common prediction network, the multiple scene prediction networks, the trained embedded feature matrix, and the weighting coefficients.

[0141] In some embodiments, the electronic device corrects the network parameters of multiple initial scene prediction networks based on the recommendation index of each target sample and the sample scene information of each sample recommendation resource to obtain multiple scene prediction networks. Specifically, based on the sample scene information of each sample recommendation resource, each target sample recommendation index is assigned to the multiple initial scene prediction networks respectively, and the network parameters of the assigned initial scene prediction networks are corrected sequentially based on the recommendation index of each target sample to obtain multiple scene prediction networks in the target prediction model.

[0142] For example, multiple initial scene prediction networks include prediction network A, prediction network B, and prediction network C. Sample scene information A in sample information set A corresponds to prediction network A, sample scene information B in sample information set B corresponds to prediction network B, and sample scene information C in sample information set C corresponds to prediction network C. The target sample recommendation index A corresponding to sample information set A is obtained based on the first sample recommendation index A and the second sample recommendation index A generated based on prediction network A. The target sample recommendation index B corresponding to sample information set B is obtained based on the first sample recommendation index B and the second sample recommendation index B generated based on prediction network B. The target sample recommendation index C corresponding to sample information set C is obtained based on the first sample recommendation index C and the second sample recommendation index C generated based on prediction network C. Therefore, the initial common prediction network can be trained using the target sample recommendation indices A, B, and C, and prediction network A can be trained using the target sample recommendation index A, prediction network B can be trained using the target sample recommendation index B, and prediction network C can be trained using the target sample recommendation index C.

[0143] In some embodiments, the electronic device can acquire recommendation index labels for each sample information set, and determine a loss value for each sample information set based on the target sample recommendation index and the recommendation index label. Therefore, correcting the network parameters of the initial common prediction network based on the recommendation index of each target sample can be done by correcting the network parameters of the initial common prediction network based on the loss value of each sample information set. Correcting the network parameters of multiple initial scene prediction networks based on the recommendation index of each target sample and the sample scene information of each sample recommendation resource can be done by correcting the network parameters of the initial scene prediction network corresponding to the sample scene information of each sample information set separately based on the loss value of each sample information set. Furthermore, the electronic device can determine a target loss value for multiple sample information sets based on the loss value of each sample information set, and determine whether the trained model has converged based on the target loss value.

[0144] For example, such as Figure 9a As shown, Figure 9a This application provides a schematic diagram of a model training scenario. The initial prediction model may include a feature generation layer, a first recommendation index generation layer, a second recommendation index generation layer, and a target recommendation index generation layer. The first recommendation index generation layer may include an initial common prediction network, and the second recommendation index generation layer may include multiple initial scene prediction networks. Each initial scene prediction network corresponds to a sample scene information. The initial scene prediction network may consist of one or more fully connected layers (FC). The number of fully connected layers in different initial common prediction networks may vary and can be set by relevant business personnel based on experience. The initial common prediction network may include an initial feature fusion network and an initial influence prediction network. The initial influence prediction network may consist of one or more fully connected layers. Wherein:

[0145] When multiple sample information sets are obtained, a sample feature set corresponding to each sample information set is generated in the feature generation layer. The sample feature set includes the corresponding encoded feature vector and embedded feature vector.

[0146] In the initial feature fusion network, the sample fusion attribute features corresponding to each sample information set are obtained based on the embedded feature vectors in each sample feature set;

[0147] In the initial influence prediction network, the sample recommendation influence value corresponding to each sample information set is obtained according to the embedded feature vector in each sample feature set. The sum of all elements in the sample fusion attribute feature corresponding to each sample information set and the sample recommendation influence value corresponding to each sample information set is determined as the first sample recommendation index corresponding to each sample information set.

[0148] Based on the initial scene prediction network corresponding to the sample scene information in each sample information set, the sample object attribute features and sample resource attribute features in each sample information set are predicted respectively, and the second sample recommendation index corresponding to each sample information set is obtained.

[0149] In the target recommendation index generation layer, the first sample recommendation index and the second sample recommendation index corresponding to each sample information set are weighted and summed to obtain the target sample recommendation index corresponding to each sample information set.

[0150] Based on the target sample recommendation index corresponding to each sample information set, the parameters of the feature generation layer, the first recommendation index generation layer, and the target recommendation index generation layer are adjusted. Based on the target sample recommendation index corresponding to each sample information set, the parameters of the initial scene prediction network corresponding to each sample information set are adjusted to obtain the target prediction model. In the application stage, if there are multiple target scene information sets, determining the second recommendation index requires using weighting coefficients for the weighted summation of the first and second initial recommendation indices. These weighting coefficients can be specified by relevant business personnel in the target prediction model, or they can be further trained based on the target prediction model, only adjusting these weighting coefficients to obtain the final target prediction model.

[0151] Optionally, a gating unit can be added to the initial prediction model to distinguish different initial scene prediction networks based on different scene information. The gating unit can determine the corresponding target scene prediction network according to the target sample scene information, input the sample object attribute features and sample resource attribute features in the target feature set into the target scene prediction network, and hide the initial scene prediction network other than the target scene prediction network from the target feature set.

[0152] It is understood that the model structure of the initial prediction model described above is merely an example and may have other forms. For instance, the model structure may differ depending on the method used to determine the recommendation index of the first sample, and / or the recommendation index of the second sample, and / or the recommendation index of the target sample. For example, the initial public prediction network may also include an initial index prediction network, which may consist of one or more fully connected layers. After obtaining the sample fusion attribute features and sample recommendation influence values, the first sample recommendation index can be determined based on this initial index prediction network. Figure 9b As shown, Figure 9b This is a schematic diagram of a prediction model provided in an embodiment of this application.

[0153] Through the above process, model prediction for multiple recommendation scenarios can be achieved. The resulting target prediction model can include a common network (common tower) for all recommendation scenarios and a unique network (unique tower) specific to a particular recommendation scenario. The final output is determined by the output of the common tower and the output of the unique tower. That is, for the target prediction model as a prediction resource, some resources are shared across multiple recommendation scenarios, while others are specific to each recommendation scenario. This allows for more comprehensive learning features incorporated into recommendation prediction, resulting in better prediction performance. During training, a batch of samples can include a set of sample information from multiple scenario scenarios, enabling scenario-based tower training within the batch data. This avoids the requirement that a batch of data contains only information from the same scenario. In other words, tower training is performed on the sample data during a batch training process, allowing for the segmentation of samples for different recommendation scenarios to train different prediction networks. Simultaneously, samples from different recommendation scenarios can be jointly trained on the common prediction network, thus fully learning the feature differences between different recommendation scenarios and fully utilizing and coordinating the features of different recommendation scenarios.

[0154] Furthermore, by jointly modeling multiple recommendation scenarios, the recommendation goals of different scenarios can be aligned. These goals can differ, allowing for an equivalent mapping of the recommendation goals for each scenario. This strengthens the correlation between different recommendation scenarios and improves the model's prediction performance. For example, ... Figure 3a The recommendation goal in the given scenario is to encourage the target audience to click on the recommended resource to generate a valid recommendation, such as... Figure 3b The recommendation goal of the current recommendation scenario is to get the target audience to watch the recommended resources for a specified period of time in order to generate effective recommendations.

[0155] For example, taking the application to live streaming recommendation tasks as an example, after a large number of tests on the target prediction model trained above, it was found that the prediction accuracy and efficiency of the target prediction model are significantly improved compared with the existing technology. Moreover, through this recommendation method, the click-through rate and viewing time of users in the live stream room have been greatly improved.

[0156] Please refer to Table 1 below:

[0157] Dataset exposure Click Duration training set 71.52 million 5.51 million 8.74 million test set 3.61 million 270,000 430,000

[0158] Table 1

[0159] The target prediction model was obtained by training the initial prediction model using the training set as the sample information set, and the target prediction model was tested using the test set. It was found that when the exposure (push notifications to the anchor) was about 3.61 million times, the number of clicks on the anchor was about 270,000 times, and the effective viewing time was about 430,000 times.

[0160] Furthermore, by comparing with existing prediction models, it was found that the target prediction model proposed in this application has a significantly improved AUC (area under the curve) compared to existing prediction models. These existing prediction models are trained using multiple batches of sample data, where each batch includes multiple sets of sample information with identical recommendation information. A larger AUC indicates better model training performance and higher prediction accuracy. See Table 2 below:

[0161] Model AUC Existing prediction models 0.8115 Target prediction model 0.8246

[0162] Table 2

[0163] Furthermore, a comparison was made with 10% of online users, and it was found that the click-through rate and viewing time obtained by applying the target prediction model proposed in this application were better than those obtained by applying the existing prediction model.

[0164] See Table 3 below:

[0165]

[0166] Table 3

[0167] Specifically, the click-through rate when using the target prediction model was approximately 1.28% higher than that when using the existing prediction model; the viewing time when using the target prediction model was approximately 2.47% higher than that when using the existing prediction model.

[0168] In this embodiment, multiple sample information sets can be obtained. Each sample information set is input into an initial prediction model. Based on the initial prediction model, a sample feature set corresponding to each sample information set is generated. The target sample recommendation index for each sample object corresponding to that sample feature set is determined based on the target sample recommendation resource. The model parameters of the initial prediction model are then adjusted based on each target sample recommendation index to obtain the target prediction model. By using the above method and jointly training multiple batches of sample data containing different sample scenario information, the target prediction model can possess prediction capabilities for multiple recommendation scenarios. Furthermore, during training, it can incorporate learning features from various recommendation scenarios, thereby improving the model's prediction efficiency and accuracy.

[0169] The methods of the embodiments of this application have been described in detail above. In order to facilitate better implementation of the above solutions of the embodiments of this application, the apparatus of the embodiments of this application is provided below.

[0170] Please see Figure 10 , Figure 10This application provides a schematic diagram of a data recommendation device; this data recommendation device can be used as a computer program (including program code) running in an electronic device, for example, the data recommendation device can be an application program in an electronic device (such as a program that can perform data recommendations). It should be noted that... Figure 10 The data recommendation device shown is used to perform the present application. Figure 2 , Figure 4 and Figure 8 Some or all of the steps in the method of the illustrated embodiment. The data recommendation device 1000 may include: an acquisition module 1001, a processing module 1002, and a determination module 1003. Wherein:

[0171] The acquisition module 1001 is used to acquire a set of recommendation information; the set of recommendation information includes object attribute information of the target object, resource attribute information of the resource to be recommended, and recommendation scenario information of the resource to be recommended.

[0172] Processing module 1002 is used to generate a set of recommendation features based on the set of recommendation information; the set of recommendation features includes object attribute features of object attribute information, resource attribute features of resource attribute information, and recommendation scenario features of recommendation scenario information;

[0173] The determination module 1003 is used to determine a first recommendation index for a target object in relation to a resource to be recommended based on a set of recommendation features; the first recommendation index represents the probability that the target object will respond to the recommendation behavior for the resource to be recommended.

[0174] The determining module 1003 is used to determine a second recommendation index for a target object in relation to a resource to be recommended under target scenario information based on object attribute features and resource attribute features; the target scenario information is at least one of multiple scenario information associated with the resource to be recommended, and the second recommendation index represents the probability that the target object responds to the recommendation behavior of the resource to be recommended;

[0175] The determination module 1003 is used to determine the target recommendation index of the target object for the resource to be recommended based on the first recommendation index and the second recommendation index; the target recommendation index represents the target probability of the target object responding to the recommendation behavior of the resource to be recommended.

[0176] The processing module 1002 is used to push the resources to be recommended to the target object's terminal according to the target recommendation index.

[0177] In one possible implementation, when determining the first recommendation index of the target object for the resource to be recommended based on the set of recommendation features, the determining module 1003 is specifically used for:

[0178] Feature fusion is performed on the recommended feature set to obtain fused attribute features;

[0179] The recommendation impact value is determined based on the set of recommendation features; this recommendation impact value characterizes the degree of influence of the set of recommendation information on the target object's response to the recommendation behavior.

[0180] The first recommendation index is determined based on the fusion attribute characteristics and the recommendation influence value.

[0181] In one possible implementation, the aforementioned multiple scene information includes recommended scene information for the resource to be recommended, and the target scene information is the recommended scene information;

[0182] When determining the second recommendation index of the target object for the resource to be recommended under the target scenario information based on object attribute characteristics and resource attribute characteristics, module 1003 is specifically used for:

[0183] Acquire multiple scene prediction networks; one scene prediction network corresponds to one scene information.

[0184] Based on the scene prediction networks corresponding to the recommended scene information in multiple scene prediction networks, the object attribute features and resource attribute features are predicted to obtain the second recommendation index.

[0185] In one possible implementation, the aforementioned multiple scene information includes recommended scene information of the resource to be recommended, and the target scene information includes recommended scene information and associated scene information of the recommended scene information among the multiple scene information;

[0186] When determining the second recommendation index of the target object for the resource to be recommended under the target scenario information based on object attribute characteristics and resource attribute characteristics, module 1003 is specifically used for:

[0187] Acquire multiple scene prediction networks; one scene prediction network corresponds to one scene information.

[0188] Based on the scene prediction networks corresponding to the recommended scene information and the scene prediction networks corresponding to the associated scene information in multiple scene prediction networks, the object attribute features and resource attribute features are predicted to obtain the first initial recommendation index of the target object for the resource to be recommended under the recommended scene information and the second initial recommendation index of the target object for the resource to be recommended under the associated scene information.

[0189] The second recommendation index is obtained by weighted summation of the first and second initial recommendation indices.

[0190] In one possible implementation, the aforementioned target recommendation index is generated by a target prediction model; the processing module 1002 is further configured to:

[0191] Obtain multiple sets of sample information; a set of sample information includes the sample object attribute information of a sample object, the sample resource attribute information of a sample recommended resource, and the sample scene information of the sample recommended resource.

[0192] Each set of sample information is input into the initial prediction model, and a set of sample features corresponding to each set of sample information is generated based on the initial prediction model. A set of sample features includes sample object attribute features of sample object attribute information, sample resource attribute features of sample resource attribute information, and sample recommendation scenario features of sample scenario information of sample resource attribute information.

[0193] Based on each sample feature set, determine the target sample recommendation index for the sample recommendation resources corresponding to each sample feature set;

[0194] The model parameters of the initial prediction model are adjusted based on the recommendation index of each target sample to obtain the target prediction model.

[0195] In one possible implementation, the initial prediction model includes an initial common prediction network and multiple initial scene prediction networks, with each initial scene prediction network corresponding to a sample scene information.

[0196] When processing module 1002 determines the target sample recommendation index for the sample object corresponding to each sample feature set based on each sample feature set, it specifically performs the following:

[0197] In the initial public prediction network, a first sample recommendation index is generated for the target sample object relative to the target sample recommendation resource based on the target feature set; the target feature set is any sample feature set in each sample feature set, the target sample object is the sample object corresponding to the target feature set, and the target sample recommendation resource is the sample recommendation resource corresponding to the target sample object;

[0198] In the target scene prediction network, a second sample recommendation index for the target sample object and the target sample resource is determined based on the sample object attribute features and the sample resource attribute features in the target feature set. The target scene prediction network is the initial scene prediction network corresponding to the target sample scene information among multiple initial scene prediction networks, and the target sample scene information is the sample scene information corresponding to the target sample recommendation resource.

[0199] The target sample recommendation index for the target sample object is determined based on the first sample recommendation index and the second sample recommendation index.

[0200] In one possible implementation, the initial prediction model includes an initial common prediction network and multiple initial scene prediction networks; when the processing module 1002 corrects the model parameters of the initial prediction model according to the recommendation index of each target sample to obtain the target prediction model, it is specifically used for:

[0201] The network parameters of the initial public prediction network are adjusted based on the recommendation index of each target sample to obtain the public prediction network.

[0202] The network parameters of multiple initial scene prediction networks are corrected based on the recommendation index of each target sample and the sample scene information of each sample recommendation resource to obtain multiple scene prediction networks.

[0203] The target prediction model is determined based on a public prediction network and multiple scenario prediction networks.

[0204] In one possible implementation, an initial scene prediction network corresponds to one sample scene information; when the processing module 1002 corrects the network parameters of multiple initial scene prediction networks based on the recommendation index of each target sample and the sample scene information of each sample recommendation resource to obtain multiple scene prediction networks, it is specifically used for:

[0205] Based on the sample scenario information of each sample's recommended resources, the recommendation index of each target sample is assigned to multiple initial scenario prediction networks.

[0206] The network parameters of the initial scene prediction network are adjusted based on the recommendation index of each target sample to obtain multiple scene prediction networks.

[0207] According to one embodiment of this application, Figure 10 The data recommendation device shown can be composed of various units, either individually or entirely, into one or more other units. Alternatively, some of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of this application. The above units are based on logical function division. In practical applications, the function of one unit can be implemented by multiple units, or the function of multiple units can be implemented by one unit.

[0208] In other embodiments of this application, the data recommendation device may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by multiple units. According to another embodiment of this application, the data recommendation device can be executed on a general-purpose computing device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM). Figure 2 , Figure 4 and Figure 8The computer program (including program code) for each step involved in the corresponding method shown, to construct such... Figure 10 The data recommendation apparatus shown herein, and the data recommendation method for implementing the embodiments of this application, are described. The computer program may be recorded on, for example, a computer-readable recording medium, loaded onto the aforementioned electronic device via the computer-readable recording medium, and run therein.

[0209] In this embodiment, the acquisition module acquires a set of recommendation information; the processing module generates a set of recommendation features based on the set of recommendation information; the determination module determines a first recommendation index for the target object relative to the resource to be recommended based on the set of recommendation features; the determination module determines a second recommendation index for the target object relative to the resource to be recommended under target scene information based on object attribute features and resource attribute features; the determination module determines a target recommendation index for the target object relative to the resource to be recommended based on the first and second recommendation indices; and the processing module pushes the resource to be recommended to the target object's terminal according to the target recommendation index. Through the above device, the obtained first recommendation index reflects the interaction of multiple recommendation features in the set of recommendation features, and the obtained second recommendation index is obtained using learned features under target scene information. These learned features under target scene information can be obtained by comprehensively learning features from all scene information. The obtained target recommendation index can not only combine its own multiple recommendation features at a deeper level but also combine learned features from multiple recommendation scenarios, which can improve the recommendation prediction efficiency and accuracy of the resource to be predicted and is beneficial to improving the effect of subsequent recommendations based on the target recommendation index.

[0210] In the various embodiments of this application, the functional modules can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules, and this application does not impose any limitations on this.

[0211] Please see Figure 11 , Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 11As shown, the electronic device 1100 includes a processor 1101, a communication interface 1102, and a computer-readable storage medium 1103. The processor 1101, communication interface 1102, and 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 is used to store computer programs, including 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 and control core of the electronic device, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve corresponding method flows or corresponding functions.

[0212] This application embodiment also provides a computer-readable storage medium (Memory), which is a memory device in an electronic device used to store programs and data. It is understood that the computer-readable storage medium here can include both built-in storage media in the electronic device and extended storage media supported by the electronic device. The computer-readable storage medium provides storage space that stores the processing system of the electronic device. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by the processor 1101, which may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here may be high-speed RAM memory or non-volatile memory, such as at least one disk storage device; optionally, it may also be at least one computer-readable storage medium located remotely from the aforementioned processor.

[0213] In one embodiment, the computer-readable storage medium stores one or more instructions; the processor 1101 loads and executes one or more instructions stored in the computer-readable storage medium to implement the corresponding steps in the above-described document processing method embodiment; specifically, the one or more instructions in the computer-readable storage medium are loaded by the processor 1101 and executed as follows:

[0214] Obtain a set of recommendation information; this set of recommendation information includes the object attribute information of the target object, the resource attribute information of the resource to be recommended, and the recommendation scenario information of the resource to be recommended.

[0215] A set of recommendation features is generated based on the set of recommendation information; this set of recommendation features includes object attribute features of object attribute information, resource attribute features of resource attribute information, and recommendation scenario features of recommendation scenario information.

[0216] The first recommendation index of the target object for the resource to be recommended is determined based on the set of recommendation features; the first recommendation index represents the probability that the target object responds to the recommendation behavior for the resource to be recommended.

[0217] The second recommendation index of the target object for the resource to be recommended is determined based on the object attribute characteristics and resource attribute characteristics; the target scenario information is at least one of multiple scenario information associated with the resource to be recommended; the second recommendation index represents the probability that the target object responds to the recommendation behavior for the resource to be recommended.

[0218] The target recommendation index for the target object is determined based on the first recommendation index and the second recommendation index, and the resource to be recommended is pushed to the target object's terminal according to the target recommendation index; the target recommendation index represents the target probability of the target object responding to the recommendation behavior of the resource to be recommended.

[0219] In one possible implementation, when determining a first recommendation index for a target object relative to a resource to be recommended based on a set of recommendation features, the processor 1101 specifically performs the following:

[0220] Feature fusion is performed on the recommended feature set to obtain fused attribute features;

[0221] The recommendation impact value is determined based on the set of recommendation features; this recommendation impact value characterizes the degree of influence of the set of recommendation information on the target object's response to the recommendation behavior.

[0222] The first recommendation index is determined based on the fusion attribute characteristics and the recommendation influence value.

[0223] In one possible implementation, the aforementioned multiple scene information includes recommended scene information for the resource to be recommended, and the target scene information is the recommended scene information;

[0224] When processor 1101 determines the second recommendation index of a target object for a resource to be recommended under the target scene information based on object attribute features and resource attribute features, it specifically performs the following:

[0225] Acquire multiple scene prediction networks; one scene prediction network corresponds to one scene information.

[0226] Based on the scene prediction networks corresponding to the recommended scene information in multiple scene prediction networks, the object attribute features and resource attribute features are predicted to obtain the second recommendation index.

[0227] In one possible implementation, the aforementioned multiple scene information includes recommended scene information of the resource to be recommended, and the target scene information includes recommended scene information and associated scene information of the recommended scene information among the multiple scene information;

[0228] When processor 1101 determines the second recommendation index of a target object for a resource to be recommended under the target scene information based on object attribute features and resource attribute features, it specifically performs the following:

[0229] Acquire multiple scene prediction networks; one scene prediction network corresponds to one scene information.

[0230] Based on the scene prediction networks corresponding to the recommended scene information and the scene prediction networks corresponding to the associated scene information in multiple scene prediction networks, the object attribute features and resource attribute features are predicted to obtain the first initial recommendation index of the target object for the resource to be recommended under the recommended scene information and the second initial recommendation index of the target object for the resource to be recommended under the associated scene information.

[0231] The second recommendation index is obtained by weighted summation of the first and second initial recommendation indices.

[0232] In one possible implementation, the aforementioned target recommendation index is generated by a target prediction model; the processor 1101 is further configured to:

[0233] Obtain multiple sets of sample information; a set of sample information includes the sample object attribute information of a sample object, the sample resource attribute information of a sample recommended resource, and the sample scene information of the sample recommended resource.

[0234] Each set of sample information is input into the initial prediction model, and a set of sample features corresponding to each set of sample information is generated based on the initial prediction model. A set of sample features includes sample object attribute features of sample object attribute information, sample resource attribute features of sample resource attribute information, and sample recommendation scenario features of sample scenario information of sample resource attribute information.

[0235] Based on each sample feature set, determine the target sample recommendation index for the sample recommendation resources corresponding to each sample feature set;

[0236] The model parameters of the initial prediction model are adjusted based on the recommendation index of each target sample to obtain the target prediction model.

[0237] In one possible implementation, the initial prediction model includes an initial common prediction network and multiple initial scene prediction networks, with each initial scene prediction network corresponding to a sample scene information.

[0238] When processor 1101 determines the target sample recommendation index for the sample object corresponding to each sample feature set based on each sample feature set, it specifically performs the following functions:

[0239] In the initial public prediction network, a first sample recommendation index is generated for the target sample object relative to the target sample recommendation resource based on the target feature set; the target feature set is any sample feature set in each sample feature set, the target sample object is the sample object corresponding to the target feature set, and the target sample recommendation resource is the sample recommendation resource corresponding to the target sample object;

[0240] In the target scene prediction network, a second sample recommendation index for the target sample object and the target sample resource is determined based on the sample object attribute features and the sample resource attribute features in the target feature set. The target scene prediction network is the initial scene prediction network corresponding to the target sample scene information among multiple initial scene prediction networks, and the target sample scene information is the sample scene information corresponding to the target sample recommendation resource.

[0241] The target sample recommendation index for the target sample object is determined based on the first sample recommendation index and the second sample recommendation index.

[0242] In one possible implementation, the initial prediction model includes an initial common prediction network and multiple initial scene prediction networks; when the processor 1101 corrects the model parameters of the initial prediction model based on the recommendation index of each target sample to obtain the target prediction model, it specifically performs the following:

[0243] The network parameters of the initial public prediction network are adjusted based on the recommendation index of each target sample to obtain the public prediction network.

[0244] The network parameters of multiple initial scene prediction networks are corrected based on the recommendation index of each target sample and the sample scene information of each sample recommendation resource to obtain multiple scene prediction networks.

[0245] The target prediction model is determined based on a public prediction network and multiple scenario prediction networks.

[0246] In one possible implementation, an initial scene prediction network corresponds to one sample scene information; when the processor 1101 corrects the network parameters of multiple initial scene prediction networks based on the recommendation index of each target sample and the sample scene information of each sample recommendation resource to obtain multiple scene prediction networks, it specifically performs the following:

[0247] Based on the sample scenario information of each sample's recommended resources, the recommendation index of each target sample is assigned to multiple initial scenario prediction networks.

[0248] The network parameters of the initial scene prediction network are adjusted based on the recommendation index of each target sample to obtain multiple scene prediction networks.

[0249] In this embodiment, the processor can acquire a set of recommendation information, generate a set of recommendation features based on the set of recommendation information, and determine a first recommendation index for the target object regarding the resource to be recommended based on the set of recommendation features. This first recommendation index represents the probability that the target object will respond to the recommendation behavior for the resource to be recommended. A second recommendation index for the target object regarding the resource to be recommended is determined based on object attribute features and resource attribute features under target scenario information. A target recommendation index for the target object regarding the resource to be recommended is determined based on the first and second recommendation indices, and the resource to be recommended is pushed to the target object's terminal according to the target recommendation index. Through this scheme, the obtained first recommendation index reflects the interaction of multiple recommendation features in the set of recommendation features, and the obtained second recommendation index is obtained using learned features under target scenario information. These learned features under target scenario information can be obtained by comprehensively learning features from all scenario information. The obtained target recommendation index can not only combine its own multiple recommendation features at a deeper level but also combine learned features from multiple recommendation scenarios, which can improve the efficiency and accuracy of recommendation prediction for the resource to be predicted and is beneficial to improving the effect of subsequent recommendations based on the target recommendation index.

[0250] This application also provides a computer program product, which includes program instructions that, when executed by a processor, can implement some or all of the steps in the above-described method.

[0251] Those skilled in the art will recognize that the units and algorithm steps of the various examples 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 implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0252] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, 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 data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0253] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this invention should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A data recommendation method, characterized in that: The method comprises: Acquire a recommendation information set; the recommendation information set includes object attribute information of the target object, resource attribute information of the resource to be recommended, and recommendation scenario 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, resource attribute features of the resource attribute information, and recommended scene features of the recommended scene information; Determine a first recommendation index of the target object for the resource to be recommended according to the recommendation feature set; the first recommendation index represents the probability of the target object responding to the recommendation behavior for the resource to be recommended; determining, according to the object attribute characteristics and the resource attribute characteristics, a second recommendation index of the target object for the resource to be recommended under target scenario information; the target scenario information is at least one scenario information among a plurality of scenario information associated with the resource to be recommended, and the second recommendation index represents the probability of the target object responding to the recommendation behavior for the resource to be recommended; Determine the target recommendation index of the target object for the resource to be recommended according to the first recommendation index and the second recommendation index, and push the resource to be recommended to the object terminal of the target object according to the target recommendation index; the target recommendation index represents the target probability of the target object responding to the recommendation behavior of the resource to be recommended; the target recommendation index is generated by a target prediction model; the target prediction model refers to the initial prediction model trained by using the target sample recommendation index of the sample object corresponding to the sample information set for the sample recommendation resource corresponding to the sample object; the target sample recommendation index corresponding to the sample information set is obtained based on the sample feature set corresponding to the sample information set; a sample information set includes sample object attribute information of a sample object, sample resource attribute information of a sample recommendation resource, and sample scene information of the sample recommendation resource; a sample feature set includes sample object attribute features of a sample object attribute information, sample resource attribute features of a sample resource attribute information, and sample recommendation scene features of the sample scene information of the sample resource attribute information.

2. The method according to claim 1, characterized in that The determining, according to the recommendation feature set, a first recommendation index of the target object for the resource to be recommended includes: Performing feature fusion on the recommended feature set to obtain fused attribute features; Determine a recommendation influence value according to the recommendation feature set; the recommendation influence value represents the influence degree of the recommendation information set on the target object's response to the recommendation behavior; The first recommendation index is determined according to the fused attribute feature and the recommendation influence value.

3. The method according to claim 1, characterized in that The multiple scene information includes recommended scene information of the resource to be recommended, and the target scene information is the recommended scene information; Determining a second recommendation index of the target object for the resource to be recommended under target scenario information according to the object attribute feature and the resource attribute feature includes: Obtain multiple scene prediction networks; one scene prediction network corresponds to one scene information; The object attribute characteristics and the resource attribute characteristics are predicted according to a scene prediction network corresponding to the recommended scene information in the multiple scene prediction networks to obtain the second recommendation index.

4. The method according to claim 1, characterized in that The multiple scene information includes recommended scene information of the to-be-recommended resource, and the target scene information includes the recommended scene information and associated scene information of the recommended scene information in the multiple scene information; Determining a second recommendation index of the target object for the resource to be recommended under target scenario information according to the object attribute feature and the resource attribute feature includes: Obtain multiple scene prediction networks; one scene prediction network corresponds to one scene information; Predicting the object attribute characteristics and the resource attribute characteristics according to the scene prediction network corresponding to the recommended scene information and the scene prediction network corresponding to the associated scene information in the multiple scene prediction networks, respectively, to obtain a first initial recommendation index of the target object for the resource to be recommended under the recommended scene information and a second initial recommendation index of the target object for the resource to be recommended under the associated scene information; A weighted sum is performed on the first initial recommendation index and the second initial recommendation index to obtain the second recommendation index.

5. The method according to claim 1, characterized in that The method further comprises: Obtain multiple sample information sets; Inputting each sample information set into an initial prediction model, and generating a sample feature set corresponding to each sample information set based on the initial prediction model; Determine, according to each sample feature set, a target sample recommendation index for a sample recommendation resource corresponding to the sample object corresponding to each sample feature set; The model parameters of the initial prediction model are corrected according to the recommendation index of each target sample to obtain the target prediction model.

6. The method according to claim 5, characterized in that The initial prediction model includes an initial public prediction network and multiple initial scene prediction networks, and one initial scene prediction network corresponds to one sample scene information; The step of respectively determining, according to each sample feature set, a target sample recommendation index for a sample recommendation resource corresponding to each sample feature set and corresponding to the sample object includes: In the initial public prediction network, a first sample recommendation index of a target sample object for a target sample recommendation resource is generated according to a target feature set; the target feature set is any one of the sample feature sets, the target sample object is a sample object corresponding to the target feature set, and the target sample recommendation resource is a sample recommendation resource corresponding to the target sample object; In a target scene prediction network, a second sample recommendation index of the target sample object for the target sample recommended resource is determined according to the sample object attribute features in the target feature set and the sample resource attribute features in the target feature set; the target scene prediction network is an initial scene prediction network corresponding to the target sample scene information in the multiple initial scene prediction networks, and the target sample scene information is sample scene information corresponding to the target sample recommended resource; Determine a target sample recommendation index of the target sample object for the target sample recommendation resource according to the first sample recommendation index and the second sample recommendation index.

7. The method according to claim 5, characterized in that The initial prediction model includes an initial public prediction network and a plurality of initial scene prediction networks; The method of correcting the model parameters of the initial prediction model according to each target sample recommendation index to obtain the target prediction model includes: According to the recommendation index of each target sample, the network parameters of the initial public prediction network are modified to obtain a public prediction network; According to the recommendation index of each target sample and the sample scene information of each sample recommendation resource, the network parameters of the multiple initial scene prediction networks are corrected to obtain multiple scene prediction networks; The target prediction model is determined according to the common prediction network and the multiple scene prediction networks.

8. The method according to claim 7, characterized in that One initial scene prediction network corresponds to one sample scene information; the network parameters of the multiple initial scene prediction networks are corrected according to the recommendation index of each target sample and the sample scene information of each sample recommendation resource to obtain multiple scene prediction networks, including: According to the sample scene information of each sample recommendation resource, respectively assigning each target sample recommendation index to the multiple initial scene prediction networks; The network parameters of the divided initial scene prediction network are corrected according to the recommendation index of each target sample to obtain the multiple scene prediction networks.

9. 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 8.

10. 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 8.

11. A computer program product, characterized in that The computer program product stores program instructions, which, when executed by a processor, cause the processor to perform the method according to any one of claims 1 to 8.

Images