An information determination method, apparatus, device, and computer-readable storage medium

Abstract

An embodiment of the present application discloses an information determination method, which includes: obtaining historical behavior information generated by a user for a target application, and determining shared feature information based on the historical behavior information; wherein the shared feature information characterizes the attributes of the user, the attributes of the behavior, and the attributes of the object; there is a correlation between the user, the behavior, and the object; determining target feature information of the target application based on the shared feature information; wherein the target feature information characterizes the basic features corresponding to the user and the basic features corresponding to the object, the correlation between the user and the object, and the cross features between the user and the object; performing model training based on the target feature information to obtain a feature model. An embodiment of the present application also discloses an information determination device, a device, and a computer-readable storage medium.

Description

Technical Field

[0001] The present application relates to the field of computer technologies, and in particular, to a method, apparatus, device, and computer-readable storage medium for information determination. Background Art

[0002] With the rapid development of e-commerce, how to recommend favorite items to users has become a key concern in the field of e-commerce. Currently, an heterogeneous graph can be constructed based on user behavior information, and a deep learning model can be used to train each node in the heterogeneous graph to obtain the feature vectors of each node, and then favorite items can be recommended to users based on the similarity between the feature vectors of each node. However, the above method significantly weakens the low-dimensional cross-feature information and has poor generalization ability. Summary of the Invention

[0003] To solve the above technical problems, embodiments of the present application are expected to provide a method, apparatus, device, and computer-readable storage medium for information determination, which solves the problems of significantly weakening low-dimensional feature information and poor generalization ability in the related art.

[0004] The technical solution of the present application is implemented as follows:

[0005] A method for information determination, the method includes:

[0006] Obtain historical behavior information generated by a user for a target application, and determine shared feature information based on the historical behavior information; wherein, the shared feature information characterizes the attributes of the user, the attributes of the behavior, and the attributes of the object; there is a correlation between the user, the behavior, and the object;

[0007] Based on the shared feature information, determine target feature information of the target application; wherein, the target feature information characterizes the basic features corresponding to the user and the basic features corresponding to the object, the correlation between the user and the object, and the cross-feature between the user and the object;

[0008] Perform model training based on the target feature information to obtain a feature model.

[0009] In the above solution, the determining the shared feature information based on the historical behavior information includes:

[0010] Analyze the historical behavior information to obtain the basic features of the user, the basic features of the behavior, and the basic features of the object; wherein, the shared feature information includes the basic features of the user, the basic features of the behavior, and the basic features of the object.

[0011] In the above solution, the determining the target feature information of the target application based on the shared feature information includes:

[0012] Construct a graph structure of the target application based on the shared feature information, and analyze the graph structure to obtain first feature information; wherein, the first feature information characterizes the relevance between the user and the object.

[0013] Based on the relationship between the shared feature information and the user, and the relationship between the shared feature information and the object, divide the shared feature information to obtain second feature information of the target application; wherein, the second feature information characterizes the basic features corresponding to the user and the basic features corresponding to the object.

[0014] Based on the second feature information, obtain third feature information of the target application; wherein, the third feature information characterizes the cross features between the user and the object, and the target feature information includes the first feature information, the second feature information, and the third feature information.

[0015] In the above solution, the constructing the graph structure of the target application based on the shared feature information includes:

[0016] Based on the shared feature information, determine a target user, a target item, and target content; wherein, the object includes the target item and the target content.

[0017] Based on the shared feature information, determine a first relevance between the target user and the target item, a second relevance between the target user and the target content, and a third relevance between the target item and the target content.

[0018] Taking the target user, the target item, and the target content as nodes, set the connection relationships between the nodes based on the first relevance, the second relevance, and the third relevance to obtain the graph structure.

[0019] In the above solution, the taking the target user, the target item, and the target content as nodes, setting the connection relationships between the nodes based on the first relevance, the second relevance, and the third relevance to obtain the graph structure includes:

[0020] Taking the target user, the target item, and the target content as nodes, set the connection relationships between the nodes based on the first relevance, the second relevance, and the third relevance to obtain an intermediate graph structure.

[0021] Optimize the nodes and the edges between the nodes in the intermediate graph structure to obtain the graph structure.

[0022] In the above solution, the analyzing the graph structure to obtain first feature information includes:

[0023] For each node in the graph structure, based on the relevance between the nodes, determine the first neighbor nodes of each node, the intermediate nodes between each node and the first neighbor nodes, and the second neighbor nodes of the intermediate nodes;

[0024] Determine the first meta-path between each node and the first neighbor nodes, and the second meta-path between the intermediate nodes and the second neighbor nodes;

[0025] Based on the first meta-path and the second meta-path, aggregate the node feature information of each node, the node feature information of the first neighbor nodes, and the node feature information of the second neighbor nodes to obtain the first feature information; wherein, the node feature information characterizes the attributes of the nodes and the relevance between the nodes and the remaining nodes in the graph structure.

[0026] In the above solution, the aggregating the node feature information of each node, the node feature information of the first neighbor nodes, and the node feature information of the second neighbor nodes based on the first meta-path and the second meta-path to obtain the first feature information includes:

[0027] Determine the first weight of each node, the second weight of the first neighbor nodes, and the third weight of the second neighbor nodes;

[0028] Based on the first weight, the second weight, the third weight, the first meta-path, and the second meta-path, aggregate the node feature information of each node, the node feature information of the first neighbor nodes, and the node feature information of the second neighbor nodes to obtain the first feature information.

[0029] In the above solution, the partitioning the shared feature information based on the relationship between the shared feature information and the user, and the relationship between the shared feature information and the object to obtain the second feature information of the target application includes:

[0030] Determine the relationship between the shared feature information and the user, and the relationship between the shared feature information and the object; wherein, the object includes items and content;

[0031] Based on the relationship between the shared feature information and the user and the relationship between the shared feature information and the object, partition the shared feature information to obtain user feature information, item feature information, and content feature information; wherein, the second feature information includes the user feature information, the item feature information, and the content feature information;

[0032] Correspondingly, obtaining the third feature information of the target application based on the second feature information includes:

[0033] Performing feature crossing on the user feature information, the item feature information, and the content feature information to obtain the third feature information.

[0034] In the above solution, training a feature model based on the target feature information includes:

[0035] Concatenating the first feature information, the second feature information, and the third feature information to obtain concatenated feature information;

[0036] Training an initial model based on the concatenated feature information to obtain the feature model.

[0037] An information determination device, the device includes:

[0038] An acquisition module, configured to acquire historical behavior information generated by a user for a target application, and determine shared feature information based on the historical behavior information; wherein, the shared feature information characterizes the attributes of the user, the attributes of the behavior, and the attributes of the object; there is a correlation between the user, the behavior, and the object;

[0039] The acquisition module is further configured to determine target feature information of the target application based on the shared feature information; wherein, the target feature information characterizes the basic features corresponding to the user and the basic features corresponding to the object, the correlation between the user and the object, and the cross features between the user and the object;

[0040] A processing module, configured to perform model training based on the target feature information to obtain a feature model.

[0041] An information determination device, the device includes: a processor, a memory, and a communication bus;

[0042] The communication bus is used to implement a communication connection between the processor and the memory;

[0043] The processor is configured to execute an information determination program in the memory to implement the steps of the above information determination method.

[0044] A computer-readable storage medium, the computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the above information determination method.

[0045] The information determination method, device, equipment, and computer-readable storage medium provided by the embodiments of the present application can obtain the historical behavior information generated by the user for the target application and determine the shared feature information based on the historical behavior information. Then, based on the shared feature information, determine the target feature information of the target application, and perform model training based on the target feature information to obtain a feature model. In this way, it is possible to determine the target feature information that represents the basic features corresponding to the user and the object, the relevance between the user and the object, and the cross-features between the user and the object, so that the feature model trained based on the target feature information not only retains the deep features of the user and the object, but also retains the basic features corresponding to the user and the object and the low-dimensional cross-features, and the generalization ability of the model is stronger, solving the problem in the related art that the low-dimensional cross-feature information is greatly weakened and the generalization ability is poor. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic flow chart of an information determination method provided by an embodiment of the present application;

[0047] Figure 2 It is a schematic flow chart of another information determination method provided by an embodiment of the present application;

[0048] Figure 3 It is a schematic flow chart of yet another information determination method provided by an embodiment of the present application;

[0049] Figure 4 It is a schematic flow chart of an information determination method provided by another embodiment of the present application;

[0050] Figure 5 It is a schematic structural diagram of an information determination system provided by an embodiment of the present application;

[0051] Figure 6 It is a schematic structural diagram of an information determination device provided by an embodiment of the present application;

[0052] Figure 7 It is a schematic structural diagram of an information determination device provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application.

[0054] It should be understood that the "embodiments of the present application" or "the foregoing embodiments" mentioned throughout the specification mean that specific features, structures, or characteristics related to the embodiments are included in at least one embodiment of the present application. Therefore, the "in the embodiments of the present application" or "in the foregoing embodiments" that appear throughout the specification do not necessarily refer to the same embodiment. In addition, these specific features, structures, or characteristics can be combined in one or more embodiments in any suitable manner. In various embodiments of the present application, the magnitude of the serial numbers of the above processes does not mean the order of execution, and the order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application. The serial numbers of the embodiments of the present application are only for description and do not represent the advantages or disadvantages of the embodiments.

[0055] Without special instructions, when the electronic device executes any step in the embodiments of the present application, it can be the processor of the electronic device that executes this step. It is also worth noting that the embodiments of the present application do not limit the order of execution of the following steps by the electronic device. In addition, the methods used to process data in different embodiments can be the same method or different methods. It should also be noted that any step in the embodiments of the present application can be independently executed by the electronic device, that is, when the electronic device executes any step in the following embodiments, it can be independent of the execution of other steps.

[0056] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0057] The embodiments of the present application provide an information determination method, which can be applied to an information determination device. Referring to Figure 1 as shown, the method includes the following steps:

[0058] Step 101, obtain historical behavior information generated by the user for the target application, and determine shared feature information based on the historical behavior information.

[0059] Among them, the shared feature information characterizes the attributes of the user, the attributes of the behavior, and the attributes of the object; there is a correlation between the user, the behavior, and the object.

[0060] In the embodiments of the present application, the target application is the application where the user performs an action. The target application can be an application in an e-commerce scenario or a sharing type of application; in a feasible manner, the target application can be a shopping platform such as a shopping website or a shopping application (Application, APP), the target application can include multiple shopping platforms or can include one shopping platform; the target application can also be a sharing platform such as a sharing website or a sharing APP. The embodiments of the present application do not limit this.

[0061] Behavior information is generated only after the user performs an action on an object. Therefore, there is a correlation among the user, the object, and the action. Historical behavior information is the behavior information generated after the user performs an action in the target application. The historical behavior information can be the information generated by the user's actions such as browsing content or purchasing items in the target application. The object refers to the object on which the action is imposed, and this object can be the item or content on which the action is imposed, etc. If the user purchases an item in the target application, then the action at this time is purchasing the item, and the object is the purchased item. If the user browses content in the target application, then the action at this time is browsing the content, and the object is the browsed content.

[0062] The attributes of the user are the information characterizing the user's features. The attributes of the user can be information such as the user identification unique code (Identity document, ID), the user's preferences, the user's browsing tendency, and the user's purchase preference. The attributes of the action are the information characterizing the action features. The attributes of the action can be information such as the type of the action, the occurrence times of the action, and the change trend of the action. The attributes of the object can be the attributes of the object on which the action is imposed. The attributes of the object are the information characterizing the features of this object. The attributes of this object can be information such as the object ID, the display mode of the object, and the copywriting of the object. If the object on which the action is imposed is an item, then the attributes of the object are also the attributes of the item. The attributes of the item can include information such as the category to which the item belongs, the cover image of the item, the description of the item, the title of the item, and the exposure times of the item. If the object on which the action is imposed is content, then the attributes of the object are also the attributes of the content. The attributes of the content can include information such as the content description, the content display mode, the content identifier, and the content layout. The shared feature information is determined based on the historical behavior information and can characterize the attributes of the user, the attributes of the action, and the attributes of the object.

[0063] In a feasible manner, information such as which items all users have clicked on this shopping platform, which content of each item they have browsed, and which items they have purchased can be obtained from a certain shopping platform, and the obtained information is used as the historical behavior information.

[0064] In the embodiments of the present application, since the historical behavior information is the behavior information generated after the user performs an action in the target application, information such as the user who performed the action, what actions were performed, and the object on which the action was imposed can be determined based on the historical behavior information. Furthermore, based on information such as the user who performed the action, what actions were performed, and the object on which the action was imposed, the attributes of the user who performed the action, the attributes of the action, and the attributes of the object on which the action was imposed can be determined, so as to obtain the shared feature information characterizing the attributes of the user, the attributes of the action, and the attributes of the object on which the action was imposed.

[0065] Step 102: Determine the target feature information of the target application based on the shared feature information.

[0066] Among them, the target feature information represents the basic features of the user, the basic features of the object, the correlation between the user and the object, and the cross features between the user and the object.

[0067] In the embodiments of the present application, the shared feature information can represent the attributes of the user, the attributes of the behavior, and the attributes of the object, and there is a certain correlation among the user, the behavior, and the object. Moreover, the user and the object are two different categories, and the feature information of the same category has a certain correlation. Therefore, based on the shared feature information, the target feature information that represents the basic features of the user, the basic features of the object, the correlation between the user and the object, and the cross features between the user and the object can be obtained. Among them, the basic features of the user and the basic features of the object are the original features of the user and the object, and the cross features between the user and the object are the low-dimensional cross features of the user and the object; and the correlation between the user and the object is the direct and indirect correlation between the user and the object, which belongs to the deep feature information of the user and the object.

[0068] Step 103: Perform model training based on the target feature information to obtain a feature model.

[0069] In the embodiments of the present application, the feature model is used to continuously optimize the feature information corresponding to the user or the feature information corresponding to the object to obtain more accurate feature information corresponding to the user or the feature information corresponding to the object. The target feature information can represent the basic features of the user, the basic features of the object, the correlation between the user and the object, and the cross features between the user and the object. The target feature information will contain some redundant information that cannot represent the user features and object features, and the information contained in the target feature information will be inaccurate. Therefore, it is necessary to perform model training based on the target feature information to obtain more accurate feature information corresponding to the user or the feature information corresponding to the object, which is convenient for subsequent recommending objects that meet the user's interests based on the feature information corresponding to the user or the feature information corresponding to the object. Moreover, the target feature information not only includes the original features of the basic features of the user and the basic features of the object, the low-dimensional cross features of the cross features between the user and the object, but also includes the deep features of the direct and indirect correlation between the user and the object. Therefore, the features in the target feature information are more comprehensive, the generalization ability of the trained feature model will be stronger, and the model accuracy will also be higher and more representative.

[0070] The information determination method provided by the embodiments of the present application obtains the historical behavior information generated by the user for the target application, determines the shared feature information based on the historical behavior information, then determines the target feature information of the target application based on the shared feature information, and performs model training based on the target feature information to obtain a feature model. In this way, the target feature information representing the basic features of the user, the basic features of the object, the relevance between the user and the object, and the cross features between the user and the object can be determined. Therefore, the feature model trained based on the target feature information not only retains the deep features of the user and the object, but also retains the basic features of the user, the basic features of the object, and the low-dimensional cross features. The generalization ability of the model is stronger, solving the problem in the related technology that the low-dimensional feature information is greatly weakened and the generalization ability is poor.

[0071] Based on the foregoing embodiments, the embodiments of the present application provide an information determination method. Referring to Figure 2 as shown, the method includes the following steps:

[0072] Step 201, the information determination device obtains the historical behavior information generated by the user for the target application, and analyzes the historical behavior information to obtain the basic features of the user, the basic features of the behavior, and the basic features of the object.

[0073] Among them, the shared feature information includes the basic features of the user, the basic features of the behavior, and the basic features of the object; the shared feature information represents the attributes of the user, the attributes of the behavior, and the attributes of the object; there is a relevance between the user, the behavior, and the object.

[0074] In the embodiments of the present application, the e-commerce platform and the sharing platform not only include items, but also include a large number of content scenarios, which may include information such as videos, pictures and texts, stores, and theme aggregation pages. These contents and items can be referred to as objects. In this case, there are two categories of objects, one is items, and the other is content. Secondly, each item has its own basic features. The basic features of the item may include information such as the category to which the item belongs, the item ID, and the brand of the item; each content also has its own basic features. The basic features of the content may include information such as the display method of the content, the material of the content, the keywords of the content, and the content ID. That is to say, each object has its own basic features. The basic features of the user may include information such as the user ID, the brand liked by the user, and the style liked by the user. The basic features of the behavior may include information such as the number of clicks, the number of purchases, and the click frequency; based on the basic features of the behavior, features such as the user browsing trend and the user purchase preference can be obtained, item statistical features such as the exposure rate, the number of purchases, and the click-through rate of the item can be obtained, and content statistical features such as the exposure rate and the click-through rate of the content can be obtained.

[0075] In the embodiments of the present application, the historical behavior information includes information such as what content the user has browsed and what items the user has purchased. Therefore, the historical behavior information can be analyzed to obtain the basic characteristics of the user, the basic characteristics of the behavior, and the basic characteristics of the object, and the obtained basic characteristics of the user, the basic characteristics of the behavior, and the basic characteristics of the object are used as shared feature information for subsequent processing of the shared feature information.

[0076] Step 202: The information determination device constructs a graph structure of the target application based on the shared feature information, and analyzes the graph structure to obtain first feature information.

[0077] The first feature information characterizes the correlation between the user and the object.

[0078] In the embodiments of the present application, the first feature information is the feature information obtained by analyzing the graph structure, which can characterize the direct or indirect correlation between the user, the content, and the item. The graph structure can represent the user behavior trajectory. In the e-commerce scenario, there are complex interlaced relationships among users, items, and content. Therefore, the correlation between users, items, and content, that is, the correlation between the user and the object, can be reflected through the graph structure.

[0079] In graph neural networks, there are two major categories: homogeneous graphs and heterogeneous graphs. Among them, the heterogeneity of heterogeneous graphs is reflected in that nodes and edges can have different types. Therefore, heterogeneous graphs have a wider range of application scenarios. Different node relationships such as user-item, user-content, and item-content can be represented in heterogeneous graphs, and different edge relationships such as attribute approximation and collaborative filtering approximation can also exist. Therefore, they have higher scalability. In one feasible way, a heterogeneous graph can be constructed based on the shared feature information to reflect the correlation between users, items, and content, and the first feature information characterizing the correlation between users, items, and content, that is, graph features, can be obtained by analyzing the heterogeneous graph.

[0080] Step 203: The information determination device divides the shared feature information based on the relationship between the shared feature information and the user, and the relationship between the shared feature information and the object, to obtain second feature information of the target application.

[0081] The second feature information characterizes the basic characteristics corresponding to the user and the basic characteristics corresponding to the object.

[0082] In the embodiments of the present application, after a user performs an action on an object, action information will be generated. Therefore, the shared feature information can be divided based on the relationship between the shared feature information and the user and the relationship between the shared feature information and the object, and the shared feature information is divided into a feature information set corresponding to the user and a feature information set corresponding to the object. Among them, since an action depends on the user or the object, the basic features of the action can be transformed into the basic features corresponding to the user and the basic features corresponding to the object. In a feasible manner, if the basic features of the action include the number of occurrences, then the feature of the number of times the user performs this action can be used as a statistical feature of the user, and the statistically obtained number of times this action is performed on each object can be used as a statistical feature of the object.

[0083] In the embodiments of the present application, based on the node types and the connection relationships between nodes in the graph structure, the relationship between the shared feature information and the user and the relationship between the shared feature information and the object can be determined. After dividing the shared feature information based on the relationship between the shared feature information and the user and the relationship between the shared feature information and the object to obtain the second feature information, only the basic features are naturally divided into different domains, and the basic features of the user, the action, and the object are retained, that is, the original features of the user and the original features of the object are retained.

[0084] Step 204: The information determination device obtains the third feature information of the target application based on the second feature information.

[0085] Among them, the third feature information represents the cross features between the user and the object. The target feature information includes the first feature information, the second feature information, and the third feature information. The target feature information represents the basic features corresponding to the user and the basic features corresponding to the object, the correlation between the user and the object, and the cross features between the user and the object.

[0086] In the embodiments of the present application, the user and the object belong to different categories, that is, different domains, and the feature information of each domain is correlated. Therefore, the third feature information representing the cross features between the user and the object can be obtained based on the second feature information.

[0087] Step 205: The information determination device splices the first feature information, the second feature information, and the third feature information to obtain the spliced feature information.

[0088] In the embodiments of the present application, the first feature information represents the relevance between the user and the object, the second feature information represents the basic features corresponding to the user and the basic features corresponding to the object, and the third feature information represents the cross features between the user and the object. The spliced feature information obtained by splicing these three together not only includes the original features of the user and the object, the low-dimensional cross features of the user and the object, but also includes the deep features of the user and the object, that is, it includes features of different dimensions. Therefore, the features of the spliced feature information are more comprehensive and have overall representativeness.

[0089] Step 206: The information determination device performs model training on the initial model based on the spliced feature information to obtain a feature model.

[0090] In the embodiments of the present application, since the spliced feature information not only includes the basic features corresponding to the user and the basic features corresponding to the object, the low-dimensional cross features of the user and the object, but also includes the deep features of the user and the object, the trained model has stronger fitting ability and generalization ability, and higher model accuracy.

[0091] In a feasible manner, a two-layer multi-head attention mechanism model can be adopted, and a residual network is added between the two-layer multi-head attention mechanism models to construct a feature interaction layer; a fully connected layer and a loss function are used to construct an output layer for result prediction; a gradient descent method is used for model training; among them, the feature interaction layer outputs high-dimensional cross features of residuals, which is an important method for accurately fitting data. The self-attention mechanism can dynamically measure the importance of graph features, basic features, and low-dimensional cross features, realize the dynamic weight balance of different features, and retain feature information of different dimensions; the fully connected layer aggregates the original features of the user and the object, the low-dimensional cross features of the user and the object, and the deep features of the user and the object, and retains the feature information of different dimensions to the greatest extent. While ensuring the model accuracy, the introduction of the original features and low-dimensional cross features retains the memory features for the model, prevents the model from overfitting, and improves the fitting ability, generalization ability, and overall representativeness of the model.

[0092] Among them, the steps of processing the spliced feature information by using the feature interaction layer can be: Let For the nth self-attention head, Let Att(K,Q,V) be the self-attention feature, then the calculation method of feature interaction is:

[0093]

[0094] Among them, X is the spliced feature information, Emb Interaction is the processed spliced feature information, K, Q, V are the inputs of the three linear layers of the self-attention mechanism, Q n 、Kn , V n are the values of Q, K, and V after linear transformation respectively, and w0, w1, and w2 are the weights of different attention heads.

[0095] In the embodiments of the present application, model training requires training samples, and positive and negative samples can be constructed based on the co-occurrence relationship between users and objects; if a user clicks on item A and item B at the same time, then item A and item B are co-clicked, which can represent the user's preference for item A and item B, or can indicate that there may be a potential interest correlation between item A and item B. Then, positive samples can be constructed based on the correlation between the user and item A and item B, and negative samples can be constructed by random sampling. Generally, in the sample combination method, the number of negative samples is much larger than that of positive samples. Since the cost of global training is too high, negative samples can be made by sampling. During the negative sampling process, negative samples can be made based on the frequency of node appearance. The higher the frequency of a node, the easier it is to be sampled as a negative sample, that is, the higher the frequency of a node, the easier it is to find a node that has no correlation with it. It should be noted that only items are used as examples here, and it can also be information such as videos or pictures.

[0096] It should be noted that for the description of the same steps and the same content in this embodiment and other embodiments, reference can be made to the description in other embodiments, and details will not be repeated here.

[0097] The information determination method provided by the embodiments of the present application can determine the target feature information representing the basic features of the user, the basic features of the object, the correlation between the user and the object, and the cross features between the user and the object, so that the feature model trained based on the target feature information not only retains the deep features of the user and the object, but also retains the basic features of the user, the basic features of the object, and the low-dimensional cross features corresponding to the user, and the generalization ability of the model is stronger, solving the problem in the related technology that the low-dimensional feature information is greatly weakened and the generalization ability is poor.

[0098] Based on the foregoing embodiments, the embodiments of the present application provide an information determination method. Refer to Figure 3 as shown, the method includes the following steps:

[0099] Step 301, the information determination device obtains the historical behavior information generated by the user for the target application, and analyzes the historical behavior information to obtain the basic features of the user, the basic features of the behavior, and the basic features of the object.

[0100] Among them, the shared feature information includes the basic features of the user, the basic features of the behavior, and the basic features of the object; the shared feature information represents the attributes of the user, the attributes of the behavior, and the attributes of the object; there is a correlation between the user, the behavior, and the object.

[0101] Step 302: The information determination device determines the target user, target item, and target content based on the shared feature information.

[0102] Among them, the object includes the first item and the first content.

[0103] In the embodiments of the present application, the target user can be all users in the target application or users who meet the first condition. The first condition can be that the number of times a user performs an action needs to exceed the first quantity threshold. At this time, the target user is the user whose number of times of performing an action exceeds the first quantity threshold; the target item can be all items in the target application or items that meet the second condition. The second condition can be that the number of times an item is clicked exceeds the second quantity threshold. At this time, the target item is the item whose number of times of being clicked exceeds the second quantity threshold; the target content can be all content in the target application. The target content can include information such as videos, pictures and texts, stores, aggregated materials, etc. The target content can also be content that meets the third condition. The third condition can be that the number of times the content is clicked exceeds the third quantity threshold. At this time, the target content is the content whose number of times of being clicked exceeds the third quantity threshold. It should be noted that the first quantity threshold, the second quantity threshold, and the third quantity threshold can be set to the same value or different values. In a feasible manner, the first quantity threshold, the second quantity threshold, and the third quantity threshold can all be set to 5; in addition, the first condition, the second condition, and the third condition can all be set according to actual business requirements, and the embodiments of the present application do not limit this.

[0104] Step 303: The information determination device determines the first relevance between the target user and the target item, the second relevance between the target user and the target content, and the third relevance between the target item and the target content based on the shared feature information.

[0105] In the embodiments of the present application, the first relevance is a direct or indirect association relationship existing between a target user and a target item; if user A has purchased the same type of items A and B multiple times, has not purchased item C, but item C belongs to the same type as items A and B, then there is a direct association relationship between user A and items A and B. The reason why user A has not purchased item C may be that item C has not appeared on the interface browsed by user A. If item C appears on the browsing interface of user A, user A has a certain probability of purchasing item C. Therefore, there is an indirect association relationship between user A and item C. The second relevance is a direct or indirect association relationship existing between a target user and target content; if user B has clicked on different types of content A and B multiple times, and has clicked on the videos of content A and B, then user B may like to watch videos. There is a direct association relationship between user B and content A and B, and there is an indirect association relationship between content A and B. The third relevance is a direct or indirect association relationship existing between a target item and target content; if within the same time period, the title of item A has been clicked multiple times, and the video of item B has been clicked multiple times, and item B and item A belong to different types, then there is an indirect association relationship between item A and item B. There is a direct association relationship between item A and its own title, and there is a direct association relationship between item B and its own video.

[0106] In a feasible manner, collaborative filtering and frequent item mining can be used to determine the first relevance, the second relevance, and the third relevance. The similarity between the feature information of the user, the feature information of the content, and the feature information of the item can also be calculated, and the edges of the graph structure are constructed through the similarity.

[0107] Step 304: The information determination device uses the target user, the target item, and the target content as nodes, and sets the connection relationship between the nodes based on the first relevance, the second relevance, and the third relevance to obtain a graph structure.

[0108] In the embodiments of the present application, if the target user is all applications in the target application, the target item is all items in the target application, and the target content is all content in the target application, then the graph structure constructed at this time is a relatively sparse graph decoupling structure, which mainly realizes a large-scale coverage of items and content. Utilizing the characteristics of nearest neighbor search in the graph structure, the overall recall rate and coverage rate are improved, and the constructed graph structure is a large-scale graph structure. In addition, after obtaining the graph structure, the graph structure can also be optimized to filter out edges and nodes with low similarity; in a feasible manner, the edges and nodes in the graph structure can be optimized by means of Bayesian probability posterior scores.

[0109] In other embodiments of the present application, if the target user is a user who meets the first condition, the target item is an item that meets the second condition, and the target content is content that meets the third condition, then step 304 can be implemented through the following steps:

[0110] Step 304a: The information determination device uses the target user, target item, and target content as nodes, and sets the connection relationships between the nodes based on the first relevance, second relevance, and third relevance to obtain an intermediate graph structure.

[0111] In the embodiments of the present application, since the target user, target item, and target content are all selected through certain conditions, an intermediate graph structure is constructed with high-frequency users, high-frequency items, and high-frequency content as nodes at this time. The constructed intermediate graph structure filters out low-frequency items, low-frequency content, and low-frequency users, and is a dense graph structure. The scale of the graph structure is small, and its main purpose is to improve the accuracy.

[0112] Step 304b: The information determination device optimizes the nodes and the edges between the nodes in the intermediate graph structure to obtain a graph structure.

[0113] In the embodiments of the present application, the graph structure is a graph structure obtained by optimizing the intermediate graph structure. In a feasible manner, a target condition can be set in advance, and then the intermediate graph structure is optimized based on the target condition; among them, the target condition can be to filter the edges with relatively low behavior-related scores and the nodes with too few covered edges to ensure that the nodes and edges in the graph structure have a relatively high frequency.

[0114] Step 305: For each node in the graph structure, the information determination device determines the first neighbor nodes of each node, the intermediate nodes between each node and the first neighbor nodes, and the second neighbor nodes of the intermediate nodes based on the relevance between the nodes.

[0115] In the embodiments of the present application, for any node in the graph structure, the first neighbor nodes are one or more nodes that are relevant to this node; it is possible that this node and the first neighbor nodes are connected by one or more nodes, and the one or more nodes connecting this node and the first neighbor nodes are the intermediate nodes; the second neighbor nodes are nodes that are relevant to the intermediate nodes. In a feasible manner, the mean aggregation method can be used to determine the k nearest neighbor nodes.

[0116] Step 306: The information determination device determines the first meta-path between each node and the first neighbor nodes, and the second meta-path between the intermediate nodes and the second neighbor nodes.

[0117] In the embodiments of the present application, for any node in the graph structure, the first meta-path is the meta-path connecting this node and the first neighbor node, and the second meta-path is the meta-path connecting the intermediate node and the second neighbor node. A meta-path is a kind of path in a heterogeneous graph and also a kind of relationship, representing prior knowledge in a graph structure. Through the meta-path, the neighboring nodes in the graph structure are connected. The utilization of graph-structured information is realized through the meta-path, and potential connections are mined. For example, in an e-commerce scenario, there are various nodes. Assuming that a user is represented as U, an item is represented as S, and content is represented as M, then the meta-path S-U-S represents the co-click relationship of items of the same user; the meta-path M-U-U-M represents the potential association of similar users' clicked content, and the meta-path S-M-U-S may represent the potential connection between the user's item click history and the items attached to the content.

[0118] Step 307: The information determination device aggregates the node feature information of each node, the node feature information of the first neighbor node, and the node feature information of the second neighbor node based on the first meta-path and the second meta-path to obtain the first feature information.

[0119] Among them, the node feature information characterizes the attributes of the node and the relevance between the node and the remaining nodes in the graph structure; the first feature information characterizes the relevance between the user and the object.

[0120] In the embodiments of the present application, through the aggregation of meta-paths, the graph structure information of the nodes in the heterogeneous graph can be completely obtained, and the fusion of node features, edge features, and graph structure information is realized. The nodes constituting the meta-path can be any two of users, items, and content, and there are 27 meta-paths in total using the meta-path aggregation method. In a feasible way, when aggregating node A, the K nearest nodes can be selected through the meta-path for clustering. If the nearest node selected is node B, the feature information after the aggregation of node A and node B is output respectively using a two-tower structure. In this way, the nodes aggregated through the meta-path contain two-level graph structure information.

[0121] Among them, step 307 can be implemented in the following way:

[0122] Step 307a: The information determination device determines the first weight of each node, the second weight of the first neighbor node, and the third weight of the second neighbor node.

[0123] In the embodiments of the present application, the first weight is the weight of each node, that is, the weight of the current node; the second weight is the weight of the first neighbor nodes associated with the current node. The second weight can be the average value of all the first neighbor nodes or the median of all the first neighbor nodes; the third weight is the weight of the intermediate nodes between the current node and the first neighbor nodes. The third weight can be the average value of all the intermediate nodes or the median of all the intermediate nodes. The embodiments of the present application do not limit the values of the first weight, the second weight, and the third weight.

[0124] Step 307b: The information determination device aggregates the node feature information of each node, the node feature information of the first neighbor nodes, and the node feature information of the second neighbor nodes based on the first weight, the second weight, the third weight, the first meta-path, and the second meta-path to obtain the first feature information.

[0125] In the embodiments of the present application, aggregating the node feature information of each node, the node feature information of the first neighbor nodes, and the node feature information of the second neighbor nodes based on the first weight, the second weight, the third weight, the first meta-path, and the second meta-path realizes the fusion of node features, edge features, and graph structure information, and realizes a clear heterogeneous node representation, that is, the first feature information representing the association between the user and the object. In a feasible manner, for the heterogeneous graph G=(V, E, T), where V is the nodes of the heterogeneous graph, E is the edges of the heterogeneous graph, and T is the type of the heterogeneous graph; the meta-path in the heterogeneous graph can be expressed as: where R N-1 is the meta-path; for any node node in the heterogeneous graph, the meta-path aggregation can be expressed as: The aggregation result after two-level aggregation of the node feature information in the graph structure using the mean aggregation method can be expressed as: topK(N(E)) means taking the K nearest neighbor nodes, δ is the aggregation relationship, and ω0, ω1, and ω2 are the weights of each node, the weights of the first neighbor nodes, and the weights of the second neighbor nodes, respectively.

[0126] Step 308: The information determination device determines the relationship between the shared feature information and the user, and the relationship between the shared feature information and the object.

[0127] Among them, the object includes items and content.

[0128] In the embodiments of the present application, determining the relationship between the shared feature information and the object is to determine the relationship between the shared feature information and the content, and the relationship between the shared feature information and the item. In a feasible manner, the relationship between the shared feature information and the user, the relationship between the shared feature information and the content, and the relationship between the shared feature information and the item can be determined based on the historical behavior information.

[0129] Step 309: The information determination device divides the shared feature information into user feature information, item feature information, and content feature information based on the relationship between the shared feature information and the user and the relationship between the shared feature information and the object.

[0130] Among them, the second feature information includes user feature information, item feature information, and content feature information; the second feature information represents the basic features corresponding to the user and the basic features corresponding to the object.

[0131] In the embodiments of the present application, the user feature information is feature information related to the user, which may include information such as user identification features, basic features of the user, statistical features of the user, and behavioral features of the user; the item feature information is feature information related to the item, which may include information such as item identification features, basic features of the item, and statistical features of the item; the content feature information is feature information related to the content, which may include information such as basic features of the content, statistical features of the content, modal features of the content, and content identification features. The shared feature information includes the basic features of the user, the basic features of the behavior, the basic features of the item, and the basic features of the content, and the behavior depends on the user, the content, and the item. Therefore, the shared feature information can be divided into user feature information, item feature information, and content feature information based on the relationship between the shared feature information and the user and the relationship between the shared feature information and the object.

[0132] Step 310: The information determination device performs feature crossing on the user feature information, item feature information, and content feature information to obtain third feature information.

[0133] Among them, the third feature information represents the cross features between the user and the object, and the target feature information includes the first feature information, the second feature information, and the third feature information; the target feature information represents the basic features corresponding to the user and the basic features corresponding to the object, the relevance between the user and the object, and the cross features between the user and the object.

[0134] In the embodiments of the present application, based on the Field-aware Factorization Machine (FFM) model, even in the case of very sparse data, reliable parameters can still be estimated for prediction. Considering the crossing between features, it models all nested variable interactions and has good performance in terms of indicators; among them, the introduction of the concept of field divides different feature fields and no longer uses a single embedding for representation, but models through more dimensional interactions, with better accuracy. Therefore, the FFM model can be used to perform feature crossing on the user feature information, item feature information, and content feature information to maximize the effect of low-dimensional cross features, prevent overfitting, and improve the generalization ability of the model.

[0135] Step 311: The information determination device splices the first feature information, the second feature information, and the third feature information to obtain spliced feature information.

[0136] Step 312: The information determination device performs model training on the initial model based on the spliced feature information to obtain a feature model.

[0137] As Figure 4 shown, in the information determination method provided by the embodiment of the present application, after obtaining shared feature information based on the historical behavior information generated by the user for the target application, a large-scale heterogeneous graph network is first constructed based on the shared feature information including the basic features of the user, the basic features of the behavior, the basic features of the content, and the basic features of the item, and then the heterogeneous graph is analyzed to obtain node feature information. First-level meta-path aggregation is performed based on each node and the first neighbor nodes of each node in the graph structure, and then second-level meta-path aggregation is performed based on the second neighbor nodes of the intermediate nodes to obtain the first feature information characterizing the relevance between the user, the content, and the item; the shared feature information is naturally divided into different domains based on different types of nodes in the heterogeneous graph to obtain user feature information, item feature information, and content feature information; feature crossing is performed on the user feature information, item feature information, and content feature information based on FFM to obtain the third feature information characterizing the low-dimensional cross features of the user, the item, and the content; then, the first feature information, the second feature information, and the third feature information are fused based on the feature interaction layer, and the prediction of the result is performed using a fully connected layer constructed based on the fully connected layer and the loss function. In addition, the feature model trained by the embodiment of the present application can be used to recommend content or items for the user. If only content recommendation is required without sorting results, the recommendation result can be directly output after calculating the feature similarity in the fully connected layer; if both content recommendation and the sorting result of the content are required, the recommendation result is output using the Sigmoid function (sigmod) after calculating the feature similarity.

[0138] It should be noted that the descriptions of the same steps and the same content in this embodiment and other embodiments can refer to the descriptions in other embodiments and will not be repeated here.

[0139] The information determination method provided by the embodiment of the present application can determine the target feature information characterizing the basic features corresponding to the user and the basic features corresponding to the object, the relevance between the user and the object, and the cross features between the user and the object, so that the feature model trained based on the target feature information not only retains the deep features of the user and the object, but also retains the basic features corresponding to the user and the basic features corresponding to the object and the low-dimensional cross features, and the generalization ability of the model is stronger, solving the problem in the related art that the low-dimensional feature information is greatly weakened and the generalization ability is poor.

[0140] In other embodiments of the present application, the information determination method provided by the embodiments of the present application can be applied to, for example, Figure 5 the information determination system shown in FIG. The system may include a shared feature unit 401, a graph node generation unit 402, a heterogeneous graph edge calculation unit 403, a co-occurrence sample construction unit 404, a heterogeneous graph element path aggregation unit 405, a domain factorization unit 406, a feature dynamic interaction unit 407, and a prediction unit 408. Among them, the shared feature unit 401 is configured to determine shared feature information representing the basic features of users, the basic features of behaviors, and the basic features of objects based on historical behavior information, and map the shared feature information to the same vector space for subsequent model training and prediction. The graph node generation unit 402 is configured to determine a target user, a target item, and a target content based on the shared feature information, and use the target user, the target item, and the target content as heterogeneous graph nodes. The heterogeneous graph edge calculation unit 403 is configured to determine a first relevance between a user and an item, a second similarity between a user and a content, and a third similarity between a content and an item based on the shared feature information, set edges between nodes based on the first relevance, the second similarity, and the third similarity, and filter the edges by a Bayesian probability posterior score method. The co-occurrence sample construction unit 404 is configured to construct training samples based on historical behavior information and the heterogeneous graph. The heterogeneous graph element path aggregation unit 405 is configured to aggregate the node feature information of each node, the node feature information of the first neighbor node, and the node feature information of the second neighbor node based on a first meta-path and a second meta-path to obtain first feature information. The domain factorization unit 406 is configured to perform feature crossing on user feature information, item feature information, and content feature information using an FFM model to obtain third feature information. The feature dynamic interaction unit 407 is configured to construct high-order crossed features based on the concatenated feature information after concatenating the first feature information, the second feature information, and the third feature information. The prediction unit 408 is configured to perform model training on an initial model based on the concatenated feature information to obtain a feature model.

[0141] Based on the foregoing embodiments, the embodiments of the present application provide an information determination device, which can be applied to the Figures 1 - 3 information determination method provided by the corresponding embodiment. Referring to Figure 6 shown in FIG., the information determination device 5 may include:

[0142] An acquisition module 51, configured to acquire historical behavior information generated by a user for a target application, and determine shared feature information based on the historical behavior information. The shared feature information represents the attributes of the user, the attributes of the behavior, and the attributes of the object. There is a relevance between the user, the behavior, and the object.

[0143] The obtaining module 51 is further configured to determine the target feature information of the target application based on the shared feature information; wherein, the target feature information characterizes the basic features of the user, the basic features of the object, the relevance between the user and the object, and the cross-features between the user and the object.

[0144] The processing module 52 is configured to perform model training based on the target feature information to obtain a feature model.

[0145] In other embodiments of the present application, the obtaining module 51 is specifically configured to perform the following steps:

[0146] Analyze the historical behavior information to obtain the basic features of the user, the basic features of the behavior, and the basic features of the object; wherein, the shared feature information includes the basic features of the user, the basic features of the behavior, and the basic features of the object.

[0147] In other embodiments of the present application, the obtaining module 51 is specifically configured to perform the following steps:

[0148] Construct a graph structure of the target application based on the shared feature information, and analyze the graph structure to obtain the first feature information; wherein, the first feature information characterizes the relevance between the user and the object.

[0149] Based on the relationship between the shared feature information and the user, and the relationship between the shared feature information and the object, divide the shared feature information to obtain the second feature information of the target application; wherein, the second feature information characterizes the basic features of the user and the basic features of the object.

[0150] Based on the second feature information, obtain the third feature information of the target application; wherein, the third feature information characterizes the cross-features between the user and the object, and the target feature information includes the first feature information, the second feature information, and the third feature information.

[0151] In other embodiments of the present application, the obtaining module 51 is specifically configured to perform the following steps:

[0152] Based on the shared feature information, determine the target user, the target item, and the target content; wherein, the object includes the target item and the target content.

[0153] Based on the shared feature information, determine the first relevance between the target user and the target item, the second relevance between the target user and the target content, and the third relevance between the target item and the target content.

[0154] Taking the target user, the target item, and the target content as nodes, set the connection relationship between the nodes based on the first relevance, the second relevance, and the third relevance to obtain a graph structure.

[0155] In other embodiments of the present application, the acquisition module 51 is specifically configured to perform the following steps:

[0156] Taking the target user, target item, and target content as nodes, set the connection relationships between the nodes based on the first relevance, second relevance, and third relevance to obtain an intermediate graph structure;

[0157] Optimize the nodes and the edges between the nodes in the intermediate graph structure to obtain a graph structure.

[0158] In other embodiments of the present application, the acquisition module 51 is specifically configured to perform the following steps:

[0159] For each node in the graph structure, based on the relevance between the nodes, determine the first neighbor nodes of each node, the intermediate nodes between each node and the first neighbor nodes, and the second neighbor nodes of the intermediate nodes;

[0160] Determine the first meta-path between each node and the first neighbor nodes, and the second meta-path between the intermediate nodes and the second neighbor nodes;

[0161] Based on the first meta-path and the second meta-path, aggregate the node feature information of each node, the node feature information of the first neighbor nodes, and the node feature information of the second neighbor nodes to obtain first feature information; wherein, the node feature information characterizes the attributes of the node and the relevance between the node and the remaining nodes in the graph structure.

[0162] In other embodiments of the present application, the acquisition module 51 is specifically configured to perform the following steps:

[0163] Determine the first weight of each node, the second weight of the first neighbor nodes, and the third weight of the second neighbor nodes;

[0164] Based on the first weight, the second weight, the third weight, the first meta-path, and the second meta-path, aggregate the node feature information of each node, the node feature information of the first neighbor nodes, and the node feature information of the second neighbor nodes to obtain first feature information.

[0165] In other embodiments of the present application, the acquisition module 51 is specifically configured to perform the following steps:

[0166] Determine the relationship between the shared feature information and the user, and the relationship between the shared feature information and the object; wherein, the object includes items and content;

[0167] Based on the relationship between the shared feature information and the user and the relationship between the shared feature information and the object, divide the shared feature information to obtain user feature information, item feature information, and content feature information; wherein, the second feature information includes user feature information, item feature information, and content feature information.

[0168] Correspondingly, the obtaining module 51 is specifically configured to perform the following steps:

[0169] Perform feature crossing on the user feature information, item feature information, and content feature information to obtain third feature information.

[0170] In other embodiments of the present application, the processing module 52 is specifically configured to perform the following steps:

[0171] Concatenate the first feature information, second feature information, and third feature information to obtain concatenated feature information;

[0172] Perform model training on the initial model based on the concatenated feature information to obtain a feature model.

[0173] It should be noted that the specific descriptions of the steps executed by each unit can be referred to Figures 1 - 3 in the information determination method provided in the corresponding embodiment, which will not be elaborated here.

[0174] The information determination device provided in the embodiments of the present application can determine target feature information characterizing the basic features corresponding to the user, the basic features corresponding to the object, the relevance between the user and the object, and the cross features between the user and the object. Therefore, the feature model trained based on the target feature information not only retains the deep features of the user and the object, but also retains the basic features corresponding to the user, the basic features corresponding to the object, and the low-dimensional cross features. The generalization ability of the model is stronger, solving the problem in the related art that the low-dimensional feature information is greatly weakened and the generalization ability is poor.

[0175] Based on the foregoing embodiments, an embodiment of the present application provides an information determination device, which can be applied to Figures 1 - 3 the information determination method provided in the corresponding embodiment, referring to Figure 7 as shown, the information determination device 6 may include: a processor 61, a memory 62, and a communication bus 63, where:

[0176] The communication bus 63 is used to implement the communication connection between the processor 61 and the memory 62;

[0177] The processor 61 is configured to execute the information determination program in the memory 62 to implement the following steps:

[0178] Obtain historical behavior information generated by the user for the target application, and determine shared feature information based on the historical behavior information; wherein, the shared feature information characterizes the attributes of the user, the attributes of the behavior, and the attributes of the object; there is a relevance between the user, the behavior, and the object;

[0179] Determine the target feature information of the target application based on the shared feature information; wherein, the target feature information characterizes the basic features of the user, the basic features of the object, the relevance between the user and the object, and the cross features between the user and the object;

[0180] Perform model training based on the target feature information to obtain a feature model.

[0181] In other embodiments of the present application, the processor 61 is used to execute the information determination program in the memory 62 to determine the shared feature information based on the historical behavior information, so as to implement the following steps:

[0182] Analyze the historical behavior information to obtain the basic features of the user, the basic features of the behavior, and the basic features of the object; wherein, the shared feature information includes the basic features of the user, the basic features of the behavior, and the basic features of the object.

[0183] In other embodiments of the present application, the processor 61 is used to execute the information determination program in the memory 62 to determine the target feature information of the target application based on the shared feature information, so as to implement the following steps:

[0184] Construct a graph structure of the target application based on the shared feature information, and analyze the graph structure to obtain the first feature information; wherein, the first feature information characterizes the relevance between the user and the object;

[0185] Based on the relationship between the shared feature information and the user, and the relationship between the shared feature information and the object, divide the shared feature information to obtain the second feature information of the target application; wherein, the second feature information characterizes the basic features of the user and the basic features of the object;

[0186] Based on the second feature information, obtain the third feature information of the target application; wherein, the third feature information characterizes the cross features between the user and the object, and the target feature information includes the first feature information, the second feature information, and the third feature information.

[0187] In other embodiments of the present application, the processor 61 is used to execute the information determination program in the memory 62 to construct a graph structure of the target application based on the shared feature information, so as to implement the following steps:

[0188] Based on the shared feature information, determine the target user, the target item, and the target content; wherein, the object includes the target item and the target content;

[0189] Based on the shared feature information, determine the first relevance between the target user and the target item, the second relevance between the target user and the target content, and the third relevance between the target item and the target content;

[0190] Taking the target user, target item, and target content as nodes, based on the first relevance, second relevance, and third relevance, set the connection relationships between the nodes to obtain a graph structure.

[0191] In other embodiments of the present application, the processor 61 is used to execute the information determination program in the memory 62, taking the target user, target item, and target content as nodes, based on the first relevance, second relevance, and third relevance, setting the connection relationships between the nodes to obtain a graph structure, so as to implement the following steps:

[0192] Taking the target user, target item, and target content as nodes, based on the first relevance, second relevance, and third relevance, set the connection relationships between the nodes to obtain an intermediate graph structure;

[0193] Optimize the nodes and the edges between the nodes in the intermediate graph structure to obtain a graph structure.

[0194] In other embodiments of the present application, the processor 61 is used to execute the information determination program in the memory 62 to analyze the graph structure to obtain first feature information, so as to implement the following steps:

[0195] For each node in the graph structure, based on the relevance between the nodes, determine the first neighbor nodes of each node, the intermediate nodes between each node and the first neighbor nodes, and the second neighbor nodes of the intermediate nodes;

[0196] Determine the first meta-path between each node and the first neighbor nodes, and the second meta-path between the intermediate nodes and the second neighbor nodes;

[0197] Based on the first meta-path and the second meta-path, aggregate the node feature information of each node, the node feature information of the first neighbor nodes, and the node feature information of the second neighbor nodes to obtain first feature information; wherein, the node feature information characterizes the attributes of the node and the relevance between the node and the remaining nodes in the graph structure.

[0198] In other embodiments of the present application, the processor 61 is used to execute the information determination program in the memory 62 to aggregate the node feature information of each node, the node feature information of the first neighbor nodes, and the node feature information of the second neighbor nodes based on the first meta-path and the second meta-path to obtain first feature information, so as to implement the following steps:

[0199] Determine the first weight of each node, the second weight of the first neighbor nodes, and the third weight of the second neighbor nodes;

[0200] Aggregate the node feature information of each node, the node feature information of the first neighbor node, and the node feature information of the second neighbor node based on the first weight, the second weight, the third weight, the first meta-path, and the second meta-path to obtain the first feature information.

[0201] In other embodiments of the present application, the processor 61 is used to execute the information determination program in the memory 62 to determine the relationship between the shared feature information and the user, and the relationship between the shared feature information and the object, and divide the shared feature information to obtain the second feature information of the target application, so as to implement the following steps:

[0202] Determine the relationship between the shared feature information and the user, and the relationship between the shared feature information and the object; wherein, the object includes items and content;

[0203] Based on the relationship between the shared feature information and the user and the relationship between the shared feature information and the object, divide the shared feature information to obtain user feature information, item feature information, and content feature information; wherein, the second feature information includes user feature information, item feature information, and content feature information.

[0204] Correspondingly, the processor 61 is used to execute the information determination program in the memory 62 to obtain the third feature information of the target application based on the second feature information, so as to implement the following steps:

[0205] Perform feature crossing on the user feature information, item feature information, and content feature information to obtain the third feature information.

[0206] In other embodiments of the present application, the processor 61 is used to execute the information determination program in the memory 62 to perform model training based on the target feature information to obtain a feature model, so as to implement the following steps:

[0207] Concatenate the first feature information, the second feature information, and the third feature information to obtain concatenated feature information;

[0208] Perform model training on the initial model based on the concatenated feature information to obtain a feature model.

[0209] It should be noted that the specific description of the steps executed by the processor can be referred to Figures 1 - 3 in the information determination method provided in the corresponding embodiment, which will not be elaborated here.

[0210] The information determination device provided by the embodiments of the present application can determine target feature information representing the basic features of the user, the basic features of the object, the relevance between the user and the object, and the cross features between the user and the object, so that the feature model trained based on the target feature information not only retains the deep features of the user and the object, but also retains the basic features of the user, the basic features of the object, and the low-dimensional cross features, and the generalization ability of the model is stronger, solving the problem in the related technology that the low-dimensional feature information is greatly weakened and the generalization ability is poor.

[0211] Based on the foregoing embodiments, an embodiment of the present application provides a computer-readable storage medium storing one or more programs, and the one or more programs can be executed by one or more processors to implement Figures 1 - 3 the steps of the information determination method provided by the corresponding embodiment.

[0212] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories and optical memories, etc.) containing computer-usable program codes.

[0213] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0214] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured product including an instruction device, and the instruction device implements the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0215] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are executed on the computer or other programmable apparatus to produce a computer-implemented process, thereby the instructions executed on the computer or other programmable apparatus provide steps for realizing the functions specified in one process or a plurality of processes and / or blocks. Figure 1 One process or a plurality of processes and / or blocks Figure 1 and steps for realizing the functions specified in one block or a plurality of blocks.

[0216] The foregoing are only preferred embodiments of the present application and are not intended to limit the scope of protection of the present application.

Claims

1. An information determination method, characterized in that, The method includes: Obtaining historical behavior information generated by a user for a target application, and determining shared feature information based on the historical behavior information; wherein, the shared feature information characterizes the attributes of the user, the attributes of the behavior, and the attributes of the object; there is a correlation between the user, the behavior, and the object; Constructing a graph structure of the target application based on the shared feature information, and analyzing the graph structure to obtain first feature information; wherein, the first feature information characterizes the correlation between the user and the object; Dividing the shared feature information based on the relationship between the shared feature information and the user, and the relationship between the shared feature information and the object, to obtain second feature information of the target application; wherein, the second feature information characterizes the basic features corresponding to the user and the basic features corresponding to the object; Based on the second feature information, obtaining third feature information of the target application; wherein, the third feature information characterizes the cross features between the user and the object; Performing model training based on the first feature information, the second feature information, and the third feature information to obtain a feature model.

2. The method according to claim 1, characterized in that, The determining the shared feature information based on the historical behavior information includes: Analyzing the historical behavior information to obtain the basic features of the user, the basic features of the behavior, and the basic features of the object; wherein, the shared feature information includes the basic features of the user, the basic features of the behavior, and the basic features of the object.

3. The method according to claim 1, characterized in that The constructing the graph structure of the target application based on the shared feature information includes: Based on the shared feature information, determining a target user, a target item, and target content; wherein, the object includes the target item and the target content; Based on the shared feature information, determining a first correlation between the target user and the target item, a second correlation between the target user and the target content, and a third correlation between the target item and the target content; Using the target user, the target item, and the target content as nodes, and setting the connection relationships between the nodes based on the first correlation, the second correlation, and the third correlation to obtain the graph structure.

4. The method according to claim 3, wherein The using the target user, the target item, and the target content as nodes, and setting the connection relationships between the nodes based on the first correlation, the second correlation, and the third correlation to obtain the graph structure includes: Using the target user, the target item, and the target content as nodes, and setting the connection relationships between the nodes based on the first correlation, the second correlation, and the third correlation to obtain an intermediate graph structure; Optimizing the nodes and the edges between the nodes in the intermediate graph structure to obtain the graph structure.

5. The method according to claim 3, characterized in that The analyzing the graph structure to obtain the first feature information includes: For each node in the graph structure, based on the correlation between the nodes, determining the first neighbor nodes of each node, the intermediate nodes between each node and the first neighbor nodes, and the second neighbor nodes of the intermediate nodes; Determine the first meta-path between each node and the first neighbor node, and the second meta-path between the intermediate node and the second neighbor node; Based on the first meta-path and the second meta-path, aggregate the node feature information of each node, the node feature information of the first neighbor node, and the node feature information of the second neighbor node to obtain the first feature information; wherein, the node feature information characterizes the attributes of the node and the relevance between the node and the remaining nodes in the graph structure.

6. The method according to claim 5, characterized in that, The aggregating the node feature information of each node, the node feature information of the first neighbor node, and the node feature information of the second neighbor node based on the first meta-path and the second meta-path to obtain the first feature information includes: Determine the first weight of each node, the second weight of the first neighbor node, and the third weight of the second neighbor node; Based on the first weight, the second weight, the third weight, the first meta-path, and the second meta-path, aggregate the node feature information of each node, the node feature information of the first neighbor node, and the node feature information of the second neighbor node to obtain the first feature information.

7. The method according to claim 1, wherein The partitioning the shared feature information based on the relationship between the shared feature information and the user, and the relationship between the shared feature information and the object to obtain the second feature information of the target application includes: Determine the relationship between the shared feature information and the user, and the relationship between the shared feature information and the object; wherein, the object includes items and content; Based on the relationship between the shared feature information and the user and the relationship between the shared feature information and the object, partition the shared feature information to obtain user feature information, item feature information, and content feature information; wherein, the second feature information includes the user feature information, the item feature information, and the content feature information; Correspondingly, the obtaining the third feature information of the target application based on the second feature information includes: Perform feature crossing on the user feature information, the item feature information, and the content feature information to obtain the third feature information.

8. The method according to claim 1, wherein The training the initial model based on the first feature information, the second feature information, and the third feature information to obtain the feature model includes: Concatenate the first feature information, the second feature information, and the third feature information to obtain concatenated feature information; Based on the concatenated feature information, train the initial model to obtain the feature model.

9. An information determination device, characterized in that, The apparatus includes: An acquisition module, configured to acquire historical behavior information generated by a user for a target application, and determine shared feature information based on the historical behavior information; wherein, the shared feature information characterizes the attributes of the user, the attributes of the behavior, and the attributes of the object; there is a relevance between the user, the behavior, and the object; The obtaining module is further configured to construct a graph structure of the target application based on the shared feature information, and analyze the graph structure to obtain first feature information; wherein the first feature information characterizes the relevance between the user and the object. The obtaining module is further configured to divide the shared feature information based on the relationship between the shared feature information and the user, and the relationship between the shared feature information and the object, to obtain second feature information of the target application; wherein the second feature information characterizes the basic features corresponding to the user and the basic features corresponding to the object. The obtaining module is further configured to obtain third feature information of the target application based on the second feature information; wherein the third feature information characterizes the cross features between the user and the object. The processing module is configured to perform model training based on the first feature information, the second feature information, and the third feature information to obtain a feature model.

10. An information determination device, characterized in that, The device includes: a processor, a memory, and a communication bus. The communication bus is used to implement a communication connection between the processor and the memory. The processor is configured to execute the information determination program in the memory to implement the steps of the information determination method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the information determination method according to any one of claims 1 to 8.

Images