Project recommendation method and device

Abstract

The embodiments of the present specification provide a project recommendation method and device, wherein the project recommendation method includes determining a target user and initial projects of different domains to be recommended to the target user; constructing a relationship feature graph based on the target user, a first associated user having a resource relationship with the target user, and a second associated user having a social relationship with the target user, wherein the first associated user and / or the second associated user belongs to the same domain or different domains as the target user; determining a target user feature of the target user and a target project feature of the initial project based on the relationship feature graph, the isomorphic graph, and the heterogeneous graph; determining a target project of the target user from the initial project based on the target user feature of the target user and the target project feature of the initial project.

Description

Technical Field

[0001] The embodiments of the present specification relate to the field of computer technology, and in particular to a project recommendation method. Background Art

[0002] Traditional recommendation problems are often trained on only one domain (which can be understood as a project line), and the recommendation model obtained from the training can only be used in this domain. However, on large industrial data sets, such recommendation models often do not work well. Since the data of some project lines is relatively sparse, there is a cold start problem. If only data from one domain is used for training, it is difficult to achieve good results.

[0003] Currently, many institutions have begun to consider feature sharing between domains. However, the recommendation model trained in this way is not suitable for fields with sparse user behavior. For example, user behavior in the cross-domain financial field is relatively sparse. If only feature sharing between domains is considered, it will make many new users with no behavior difficult to train. Then, for these new users, the recommendation model will not be able to give good recommendation results. Summary of the invention

[0004] In view of this, the present specification provides a project recommendation method. One or more embodiments of the present specification also relate to a project recommendation device, a computing device, a computer-readable storage medium, and a computer program to solve the technical defects existing in the prior art.

[0005] According to a first aspect of an embodiment of this specification, a project recommendation method is provided, including:

[0006] Determine a target user and initial items of different domains to be recommended to the target user;

[0007] Building a relationship feature graph based on the target user, a first associated user having a resource relationship with the target user, and a second associated user having a social relationship with the target user, wherein the first associated user and / or the second associated user and the target user belong to the same domain or different domains;

[0008] Determine the target user feature of the target user and the target project feature of the initial project based on the relationship feature graph, the isomorphic graph, and the heterogeneous graph;

[0009] Based on the target user characteristics of the target user and the target item characteristics of the initial items, the target item of the target user is determined from the initial items.

[0010] According to a second aspect of an embodiment of this specification, there is provided an item recommendation device, including:

[0011] a target user determination module, configured to determine a target user and initial items of different domains to be recommended to the target user;

[0012] A feature graph construction module is configured to construct a relationship feature graph based on the target user, a first associated user having a resource relationship with the target user, and a second associated user having a social relationship with the target user, wherein the first associated user and / or the second associated user belongs to the same domain or different domains as the target user;

[0013] A feature determination module, configured to determine a target user feature of the target user and a target project feature of the initial project based on the relationship feature graph, the isomorphic graph, and the heterogeneous graph;

[0014] The project recommendation module is configured to determine a target project of the target user from the initial projects based on a target user feature of the target user and a target project feature of the initial project.

[0015] According to a third aspect of an embodiment of this specification, a computing device is provided, including:

[0016] Memory and processor;

[0017] The memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions. When the computer executable instructions are executed by the processor, the steps of the above-mentioned project recommendation method are implemented.

[0018] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided, which stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the steps of the above-mentioned project recommendation method are implemented.

[0019] According to a fifth aspect of the embodiments of this specification, a computer program is provided, wherein when the computer program is executed in a computer, the computer is caused to execute the steps of the item recommendation method described in the claims.

[0020] An embodiment of the present specification implements a project recommendation method and device, wherein the project recommendation method includes determining a target user and initial projects of different domains to be recommended to the target user; constructing a relationship feature graph based on the target user, a first associated user having a resource relationship with the target user, and a second associated user having a social relationship with the target user, wherein the first associated user and / or the second associated user belongs to the same domain or different domains as the target user; determining a target user feature of the target user and a target project feature of the initial project based on the relationship feature graph, the isomorphic graph, and the heterogeneous graph; and determining a target project of the target user from the initial project based on the target user feature and the target project feature of the initial project.

[0021] Specifically, the project recommendation method captures information between users through a relationship feature graph that includes the resource relationships and social relationships of the target users, thereby better solving the cold start problem; and the user-side features can be enhanced through the resource relationships and social relationships of the target users in the same domain or different domains, thereby greatly enhancing the user features of the target users and avoiding the problem of being unable to make good project recommendations for new users by only considering feature sharing between domains. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a flow chart of a project recommendation method provided by an embodiment of this specification;

[0023] Figure 2 It is a schematic diagram of training a recommendation model in an item recommendation method provided in one embodiment of this specification;

[0024] Figure 3 It is a structural schematic diagram of a project recommendation device provided by an embodiment of this specification;

[0025] Figure 4 It is a structural block diagram of a computing device provided by an embodiment of this specification. DETAILED DESCRIPTION

[0026] Many specific details are described in the following description to facilitate a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of this specification, so this specification is not limited to the specific implementation disclosed below.

[0027] The terms used in one or more embodiments of this specification are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of this specification. The singular forms of "a", "said" and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0028] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of this specification, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0029] First, the terms involved in one or more embodiments of this specification are explained.

[0030] Graph: A graph is a structure that describes a set of objects, where some objects are "related" in some sense. These objects are called vertices (also called nodes or points), and the abstract relationship between each related vertex is called an edge (also called a link or line).

[0031] Graph Neural Networks: GNNs for short; is a neural network that can process graph structured data. It aggregates or transfers information about the neighborhood of the target node, thereby updating the features on the node and performing downstream tasks.

[0032] Multi-edge graph: A multi-edge graph is a graph neural network with multiple relationship edges. Each relationship describes a different connection method between nodes, and different relationships are mapped to different spaces.

[0033] Message Passing: Message Passing refers to the process of information interaction between nodes in a graph neural network. After this process, the node information will be updated.

[0034] Embedding: A common layer in deep learning network models, mainly used to process vector representations of sparse features; it not only solves the length problem of one-hot vectors, but also represents the similarity between features.

[0035] Cross-domain Recommendation: Cross-domain Recommendation refers to using data from different domains to train models. The model can solve the problem of multiple domain recommendations.

[0036] Social relationships: including but not limited to family relationships, classmate relationships, teacher-student relationships, social function relationships, emotional relationships, social relationships (such as address book friends, instant messaging tool friends), etc.

[0037] In order to solve the above technical problems of cross-domain recommendation, feature migration or collaborative filtering methods can be used. Specifically, the specific implementation methods of feature migration and collaborative filtering are as follows:

[0038] Feature transfer: Feature training is performed for each scenario (which can be understood as a project line). In the middle part of the network, the features of each scenario are transferred to the target domain to strengthen the features of the target domain. In different scenarios, feature transfer is prone to negative transfer. Because there is no setting for how to avoid negative transfer for specific scenarios, the recommendation effect will decrease after feature transfer.

[0039] Collaborative filtering: Use the behavioral relationship between users and items to build a matrix for similar users, and then use the matrix to predict items that have not been exposed. Different collaborative filtering methods are often divided into two cases, one is to use graph structure to predict collaborative filtering, and the other is non-graph. However, only combining the relationship between items to perform collaborative filtering will cause problems in financial scenarios. For example, since the behavior of users in financial scenarios is low-frequency and long-term, the user's behavior data is often sparse. When performing collaborative filtering on sparse users, there will inevitably be deviations, resulting in a decrease in recommendation effect.

[0040] Based on this, in this specification, a project recommendation method is provided. This specification also involves a project recommendation device, a computing device, a computer-readable storage medium and a computer program, which are described in detail one by one in the following embodiments.

[0041] See also Figure 1 , Figure 1 A flowchart of a project recommendation method provided according to an embodiment of the present specification is shown, which specifically includes the following steps.

[0042] Step 102: Determine a target user and initial items of different domains to be recommended to the target user.

[0043] Specifically, the project recommendation method provided in the embodiments of this specification can be applied to the cross-domain field of finance, such as recommending financial projects; it can also be applied to other cross-domain fields, such as the cross-domain field of entertainment, to recommend TV series, movies or music.

[0044] In actual applications, the specific application fields of the project recommendation method are different, and its target users are also different. For example, when the project recommendation method is applied to the financial cross-domain field, the target users can be understood as users with capital needs; while when the project recommendation method is applied to the entertainment cross-domain field, the target users can be understood as users with entertainment needs.

[0045] During specific implementation, determining the target user and the initial projects of different domains to be recommended to the target user can be understood as obtaining the target user and one, two or more initial projects belonging to different domains to be recommended to the target user; when the project recommendation method provided in the embodiment of this specification is applied to a financial cross-domain scenario, the initial project can be understood as any kind of financial project, such as a short-term loan project, a long-term loan project or other investment projects.

[0046] In an achievable embodiment, after determining the target user and the initial items of different domains to be recommended to the target user, the target user and the initial items can be input into a pre-trained recommendation model, and the probability value of the target user clicking on each initial item can be obtained through the pre-trained recommendation model, so as to select a better target item that the user is more interested in from the initial items based on the probability value. The specific implementation method is as follows:

[0047] After determining the target user and the initial items of different domains to be recommended to the target user, the method further includes:

[0048] Inputting the target user and the initial item into a recommendation model to obtain a probability value of the initial item relative to the target user;

[0049] A target item for the target user is determined based on the probability value.

[0050] The probability value of the initial item relative to the target user can be understood as the probability value of the target user clicking on each initial item, and the interest degree of the target user in each initial item can be determined through the probability value.

[0051] In addition, the probability value of the initial item relative to the target user can also be understood as the conversion rate of the initial item relative to the target user, through which the target user's purchase interest in each initial item can be determined. For ease of understanding, the embodiments of this specification are described in detail by taking the probability value of the initial item relative to the target user as the probability value of the target user clicking on each initial item as an example.

[0052] Before using the pre-trained recommendation model to obtain the probability value of the initial project relative to the target user and determine the target project of the target user based on the probability value, the recommendation model needs to be pre-trained to obtain the trained recommendation model for subsequent practical applications. The specific implementation method is as follows:

[0053] The recommendation model is trained through the following steps:

[0054] determining a sample user, sample items of different domains recommended to the sample user, and a sample tag of the sample user relative to the sample item;

[0055] Building a sample relationship feature graph based on the sample user, a first sample user having a resource relationship with the sample user, and a second sample user having a social relationship with the sample user;

[0056] Determine the target sample user features of the sample user and the target sample item features of the sample item based on the sample relationship feature graph, the isomorphic graph, and the heterogeneous graph;

[0057] Inputting the target sample user features of the sample user and the target sample item features of the sample item into the prediction layer of the recommendation model to obtain the probability value of the sample item relative to the sample user;

[0058] A loss function of the recommendation model is calculated based on the probability value and the sample label, and the recommendation model is trained based on the loss function.

[0059] See also Figure 2 , Figure 2 A schematic diagram of training a recommendation model in an item recommendation method provided according to an embodiment of the present specification is shown.

[0060] Figure 2 It includes four parts: sample relationship feature graph, isomorphic graph, heterogeneous graph and prediction layer. The isomorphic graph includes three parts: relationship information fusion, user information fusion and domain information fusion. The following is a detailed description of the construction of sample relationship feature graph, the role of isomorphic graph, heterogeneous graph and prediction layer in combination with actual application scenarios.

[0061] and, Figure 2Where u represents the sample user, d represents the domain, r represents the relationship between sample users or between sample users and domains, and i represents the sample item; represents the features of U at the Lth layer of domain 1, Indicates the probability predicted by the recommendation model, BCE = BCE loss, a loss function determined based on the probability of the recommendation model. The above is only an example of a specific parameter, and the explanations of other parameters of the same type are as above.

[0062] Still taking the project recommendation method provided in the embodiments of this specification applied to the financial cross-domain scenario as an example, sample users can be understood as users to whom financial projects have been recommended in the past; sample projects can be understood as financial projects in different domains recommended to sample users. For example, in a certain financial application, funds belong to one domain, financial management belongs to one domain, and insurance belongs to one domain, and each domain contains at least one financial project; the sample label of the sample user relative to the sample project can be understood as whether the sample user clicks on or purchases a certain sample project. For example, when the sample label of the sample user relative to the sample project is 1, it can be understood that the sample user clicked on or purchased a certain sample project; when the sample label of the sample user relative to the sample project is 0, it can be understood that the sample user did not click on or purchase a certain sample project.

[0063] After determining the sample user, sample project and sample label, obtain the first sample user who has a resource relationship with the sample user and the second sample user who has a social relationship with the sample user, so as to construct a relationship feature graph based on the sample user, the first sample user, the second sample user, the resource relationship between the sample user and the first sample user, and the social relationship between the sample user and the second sample user. Among them, the resource relationship can be understood as a financial relationship, such as a transfer, payment, loan and other relationships; the social relationship can be understood as a social relationship, such as a friend relationship, a kinship relationship, a superior-subordinate relationship, etc.

[0064] In addition, the first sample user who has a resource relationship with the sample user can be understood as the first sample user in the same domain or different domains who has a resource relationship with the sample user; the second sample user who has a resource relationship with the sample user can be understood as the second sample user in the same domain or different domains who has a social relationship with the sample user.

[0065] Before constructing the sample relationship feature graph, in order to enhance the information expression capability of the graph structure, the dense features of the capital relationship network and the social relationship network can be extracted in the feature extraction layer of the recommendation model, and then the extracted features can be spliced ​​and embedded in the embedding layer of the recommendation model to form a sample relationship feature graph. Among them, the capital relationship network can be understood as a polygonal graph constructed by the sample user, the first sample user, and the capital relationship between the sample user and the first sample user; the social relationship network can be understood as a polygonal graph constructed by the sample user, the second sample user, and the social relationship between the sample user and the second sample user. The specific implementation method is as follows:

[0066] Step 1: Take the sample user and the first sample user in the same domain or different domain who has a financial relationship with the sample user as nodes, and take the financial relationship between the sample user and the first sample user as an edge to build a financial relationship network;

[0067] Step 2: Take the sample user and the second sample user in the same domain or different domain who has a social relationship with the sample user as nodes, and take the social relationship between the sample user and the second sample user as edges to build a social relationship network;

[0068] Step 3: Extract dense features from the capital relationship network through the feature extraction layer of the recommendation model, and extract dense features from the social relationship network through the feature extraction layer of the recommendation model;

[0069] Specifically, performing dense feature extraction on the capital relationship network and performing dense feature extraction on the social relationship network can be understood as performing column normalization on the capital relationship network and the social relationship network respectively.

[0070] Step 4: Concatenate the dense class features of the capital relationship network extracted by the feature extraction layer with the dense class features of the social relationship network;

[0071] Step 5: Use a matrix to embed the dense features of the concatenated capital relationship network and the dense features of the social relationship network to obtain the embedded sample relationship feature graph.

[0072] Specifically, the dense features of the two networks are embedded to ensure that the numerical features provided by the two networks are in the same space.

[0073] In the sample relationship feature graph, for a node (i.e., user), its neighbor nodes and adjacent edge features are aggregated to update the node. For an edge (relationships between users, relationships between users and domains, etc.), the nodes involved are aggregated and the features of the nodes are used to update the edge.

[0074] After the sample relationship feature graph is constructed, the sample relationship feature graph is passed through the homogeneous graph network layer and the heterogeneous graph network layer in the recommendation model to obtain the target sample user features of the sample user and the target sample project features of the sample project. That is, the information transmission of the multi-edge graph, the information transmission within the domain, and the information transmission between domains are realized according to the homogeneous graph network layer in the recommendation model of the sample relationship feature graph. The specific implementation method is as follows:

[0075] Step 1: Determine the features of each edge from the sample relationship feature graph, transfer the information between each edge associated with the sample user and its adjacent edges in the isomorphic graph network layer of the recommendation model, transfer the cross features to the node of the sample user, and obtain the multi-edge graph features of the sample user;

[0076] Step 2: Determine the user features of the first sample user and / or the second sample user who are adjacent to the sample user and belong to the same domain from the sample relationship feature graph, transfer the user features in the isomorphic graph network layer of the recommendation model, transfer the domain features to the node of the sample user, and obtain the domain features of the sample user;

[0077] Specifically, for a user and its adjacent users in the same domain, the designed attention mechanism structure will be used to take the target user's features as the query and the source user's features as the key and value. After using the query and key to obtain the attention factor, the value is then used for attention, and information is transferred through this attention mechanism.

[0078] Step 3: Determine the domain features of different domains associated with the sample user from the sample relationship feature graph, transfer the domain features in the isomorphic graph network layer of the recommendation model, transfer the inter-domain features to the node of the sample user, and obtain the inter-domain features of the sample user;

[0079] Specifically, for the features of a user in each domain, the features of other domains are used as keys and values, and the features of the current domain are used as queries to perform an attention fusion, and information is transmitted through this attention mechanism.

[0080] In practical applications, the purpose of using attention to transfer information in homogeneous graphs within or between domains is to concretize the granularity of the weight parameters of information transfer to the user dimension, so that user features can learn more useful features. The specific implementation is equivalent to the attention mechanism.

[0081] Step 4: Fuse the polygonal graph features, intra-domain features, and inter-domain features of the sample user to obtain the fused features of the sample user.

[0082] After obtaining the fused features of the sample user through the homogeneous graph network layer of the recommendation model, the fused features of the sample user and the initial project features of the sample project are input into the heterogeneous graph network layer of the recommendation model. Through the heterogeneous graph network layer, the target sample user features of the sample user and the target sample project features of the sample project are obtained.

[0083] In practical applications, the heterogeneous graph network layer will perform conventional idembedding processing on the initial project features of the sample users, converting the id-class features into dense-class features, that is, the target sample project features of the sample projects. And the homogeneous graph and the heterogeneous graph will be superimposed with L layers (the output of each layer can be used as the input of the next layer), and the target sample user features of the sample users and the target sample project features of the sample projects will be output.

[0084] Finally, the target sample user features of the sample users and the target sample item features of the sample items are input into the prediction network of the recommendation model to obtain the probability value of the sample user clicking or purchasing each sample item. Finally, the loss function of the recommendation model is calculated based on the probability value and the sample label. The recommendation model is trained based on the loss function to obtain the trained recommendation model.

[0085] Step 104: construct a relationship feature graph based on the target user, a first associated user having a resource relationship with the target user, and a second associated user having a social relationship with the target user.

[0086] The first associated user and / or the second associated user and the target user belong to the same domain or different domains.

[0087] Specifically, the first associated user can be understood as a user who belongs to the same domain as the target user and has a resource relationship, and a user who belongs to a different domain than the target user and has a resource relationship; the second associated user can be understood as a user who belongs to the same domain as the target user and has a social relationship, and a user who belongs to a different domain than the target user and has a social relationship.

[0088] In a specific implementation, the building of a relationship feature graph based on the target user, a first associated user having a resource relationship with the target user, and a second associated user having a social relationship with the target user includes:

[0089] Based on the target user, determining a first associated user having a resource relationship with the target user and a second associated user having a social relationship with the target user;

[0090] constructing a first multi-edge graph and a second multi-edge graph based on the target user, the first associated user, the second associated user, the resource relationship between the target user and the first associated user, and the social relationship between the target user and the second associated user;

[0091] A relationship feature graph is constructed based on the first polygon graph and the second polygon graph.

[0092] Among them, the first polygonal graph can be understood as the financial relationship network in the above embodiment, the second polygonal graph can be understood as the social relationship network in the above embodiment, and the specific construction method of the relationship feature graph can refer to the construction of the sample relationship feature graph in the above embodiment.

[0093] In actual applications, first, based on the target user, multiple first associated users who have a financial relationship with the target user, in the same domain or different domains, and multiple second associated users who have a social relationship with the target user, in the same domain or different domains, are obtained; then, a first multilateral graph (fund relationship network) and a second multilateral graph (social relationship network) are constructed according to the target user, the first associated users, the second associated users, the fund relationship between the target user and the first associated users, and the social relationship between the target user and the second associated users; finally, a relationship feature graph is constructed based on the fund relationship network and the social relationship network.

[0094] The target recommendation method provided in the embodiments of this specification constructs a relationship feature graph based on the financial relationship network and the social relationship network, and strengthens the user-side features by introducing the financial relationship network and the social relationship network to improve the accuracy of subsequently determining target projects for target users.

[0095] Each multi-sided graph is composed of nodes and edges. Therefore, in the embodiment of this specification, in order to effectively and accurately determine the target project of the target user in combination with the graph, it is necessary to first construct a multi-sided graph based on the target user, the first associated user, the second associated user, the resource relationship, and the social relationship. The specific implementation method is as follows:

[0096] The constructing of a first multi-edge graph and a second multi-edge graph based on the target user, the first associated user, the second associated user, the resource relationship between the target user and the first associated user, and the social relationship between the target user and the second associated user includes:

[0097] Taking the target user and the first associated user as nodes and the resource relationship between the target user and the first associated user as edges, constructing a first multi-edge graph;

[0098] A second multi-edge graph is constructed by taking the target user and the second associated user as nodes and taking the social relationship between the target user and the second associated user as edges.

[0099] Among them, the multilateral graph is a graph neural network with multiple relationship edges. Each relationship describes a different connection method between nodes, and different relationships are mapped to different spaces.

[0100] Specifically, the target user and each first associated user are used as nodes, and the financial relationship between the target user and each first associated user is used as an edge to construct a first polygonal graph; the target user and each second associated user are used as nodes, and the social relationship between the target user and each second associated user is used as an edge to construct a second polygonal graph. Subsequently, a relationship feature graph can be constructed based on the first polygonal graph and the second polygonal graph, and the features of the financial relationship network and the social relationship network can be applied to the project recommendation method provided in the embodiments of this specification by means of a graph.

[0101] In practical applications, different relationships (edges) in each polygonal graph are mapped to different spaces. In order to ensure that the numerical features provided by the first polygonal graph and the second polygonal graph are in one space and their features can be applied later, the features of the two polygonal graphs need to be embedded. The specific implementation method is as follows:

[0102] The constructing a relationship feature graph based on the first polygon graph and the second polygon graph includes:

[0103] Extracting edge features of each edge in the first polygonal graph and edge features of each edge in the second polygonal graph;

[0104] Performing feature splicing on the edge feature of each edge in the first polygonal graph and the edge feature of each edge in the second polygonal graph to form an initial feature graph;

[0105] Based on a preset matrix, feature embedding is performed on the edge features of each edge in the initial feature graph to obtain a relationship feature graph.

[0106] Among them, the preset matrix can be set according to the actual application, and the embodiments of this specification do not impose any limitation on this. Any matrix that can realize the space conversion function is available.

[0107] In the specific implementation, first, the dense class features of each edge in the first polygonal graph and the dense class features of each edge in the second polygonal graph are extracted respectively (that is, the dense class features of the two polygonal graphs are column normalized respectively); then the dense class features of each edge in the first polygonal graph and the dense class features of each edge in the second polygonal graph are spliced, and finally a preset matrix is ​​used to embed the dense class features of the edges in the initial feature graph formed after splicing (that is, the feature space is transformed) to obtain a relationship feature graph.

[0108] In the embodiments of this specification, a preset matrix is ​​used to embed the dense features of the spliced ​​edges to achieve feature space transformation of each edge in the initial feature graph, enhance the information expression capability of the graph structure (relation feature graph), and subsequently accurately determine the target project of the target user based on the relationship feature graph.

[0109] Step 106: Determine the target user features of the target user and the target project features of the initial project based on the relationship feature graph, the isomorphic graph, and the heterogeneous graph.

[0110] After obtaining the relationship feature graph, we can use the information transmission method of two different domains in the graph, and combine the relationship information to transmit the information together, and then use the attention structure to achieve more fine-grained transmission between different nodes, so as to achieve adaptive information transmission for different nodes and different domains, thereby solving the problem of negative transfer. The specific implementation method is as follows:

[0111] The determining the target user feature of the target user and the target project feature of the initial project based on the relationship feature graph, the isomorphic graph, and the heterogeneous graph includes:

[0112] Determine, based on the relationship feature graph, edge features of the edge associated with the target user, user features of users in the same domain associated with the target user, and domain features of different domains associated with the target user;

[0113] The edge feature, the user feature, and the domain feature are transmitted through a homogeneous graph to obtain a fused feature of the target user;

[0114] The fusion features of the target user and the initial project features of the initial project are input into a heterogeneous graph to obtain the target user features of the target user and the target project features of the initial project.

[0115] Among them, the user characteristics of users in the same domain associated with the target user and the domain characteristics of different domains associated with the target user can be understood as the information of the same domain and different domains in the graph; the edge characteristics of the edges associated with the target user can be understood as relationship information.

[0116] Specifically, obtain the user characteristics of users in the same domain associated with the target user, the domain characteristics of different domains associated with the target user, and the edge characteristics of the edges associated with the target user from the relationship feature graph; then, pass the user characteristics of users in the same domain associated with the target user, the domain characteristics of different domains associated with the target user, and the edge characteristics of the edges associated with the target user through an isomorphic graph for information transmission to obtain the fusion characteristics of the target user; finally, input the fusion characteristics of the target user and the initial item characteristics of the initial item into the heterogeneous graph to obtain the target user characteristics of the target user and the target item characteristics of the initial item.

[0117] In specific implementation, the step of passing the edge characteristics, the user characteristics, and the domain characteristics through an isomorphic graph for information transmission to obtain the fusion characteristics of the target user includes:

[0118] Pass the edge characteristics, the user characteristics, and the domain characteristics through an isomorphic graph for information transmission to obtain the first characteristic, the second characteristic, and the third characteristic of the target user;

[0119] Fuse the first characteristic, the second characteristic, and the third characteristic to obtain the fusion characteristics of the target user.

[0120] Among them, the edge characteristics, the user characteristics, and the domain characteristics all perform information transmission through attention in the isomorphic graph. The purpose is to make the granularity of the weight parameters of information transmission concrete to the user dimension, so that the user characteristics can learn more useful characteristics.

[0121] After passing the edge characteristics, the user characteristics, and the domain characteristics through an isomorphic graph for information transmission to obtain the first characteristic, the second characteristic, and the third characteristic of the target user, fuse the first characteristic, the second characteristic, and the third characteristic to obtain the fusion characteristics of the target user.

[0122] In the item recommendation method provided by the embodiments of this specification, a domain - adaptive graph information transmission structure is used, including intra - domain graph information transmission, which can transmit information within the same domain among different users (that is, user characteristics are transmitted through an isomorphic graph); inter - domain graph information transmission, which is completed through the attention mechanism to adaptively transmit information of different domains of the same user (that is, domain characteristics are transmitted through an isomorphic graph), and combines relationship information to transmit information together (that is, edge characteristics are transmitted through an isomorphic graph), which greatly solves the problem of negative transfer of inter - domain information.

[0123] Then the edge features, user features and domain features transmit information through the isomorphic graph, and the specific implementation methods of obtaining the first feature, the second feature and the third feature are as follows:

[0124] The step of transmitting information among the edge feature, the user feature, and the domain feature through an isomorphic graph to obtain the first feature, the second feature, and the third feature of the target user includes:

[0125] Transferring information of the edge feature through a homogeneous graph to obtain a first feature of the edge associated with the target user transferred to the target user;

[0126] The user feature is transmitted through the isomorphic graph to obtain a second feature transmitted to the target user by the user in the same domain associated with the target user;

[0127] The domain features are transmitted through the isomorphic graph to obtain a third feature transmitted from the different domains associated with the target user to the target user.

[0128] In practical applications, the first feature can be understood as the polygon feature of the above embodiment, the second feature can be understood as the intra-domain feature of the above embodiment, and the third feature can be understood as the inter-domain feature of the above embodiment.

[0129] Specifically, the specific implementation process of transmitting edge features, user features and domain features through isomorphic graphs to obtain the first features, second features and third features of the target user can refer to the specific implementation methods of the multilateral graph features, intra-domain features and inter-domain features of the sample users in the above embodiments, which will not be repeated here.

[0130] In the embodiments of this specification, compared with the above-mentioned feature transfer method, the project recommendation method of the embodiments of this specification designs an information transmission method of two different domains in the graph, and combines the relationship information to transmit the information together, and uses the attention structure to achieve a more fine-grained transmission between different nodes, so as to achieve adaptive information transmission for different nodes and different domains, thereby solving the problem of negative transfer and improving the recommendation effect of subsequent target projects.

[0131] In addition, after the first feature, the second feature, and the third feature are fused to obtain the fused feature of the target user, the initial project feature of the initial project is extracted, and then the fused feature of the target user and the initial project feature of the initial project are input into the heterogeneous graph to obtain the target feature of the target user and the target project feature of the initial project.

[0132] Step 108: Determine the target project of the target user from the initial projects based on the target user characteristics of the target user and the target project characteristics of the initial projects.

[0133] In actual applications, the heterogeneous graph does not process the fused features of the target user, but only takes the fused features of the target user as input to output the target features of the target user; the heterogeneous graph performs conventional id embedding processing on the initial project features of the initial user, converting the id class features into dense class features, that is, the target sample project features of the initial project; and after obtaining the target user features of the target user and the target project features of the initial project, the probability value of the target user relative to each initial project can be determined based on the pre-trained prediction model, so as to select the target project matching the target user based on the probability value. The specific implementation method is as follows:

[0134] The determining the target project of the target user from the initial project based on the target user characteristics of the target user and the target project characteristics of the initial project includes:

[0135] Inputting the target user characteristics of the target user and the target item characteristics of the initial item into a prediction model to obtain a probability value of the initial item relative to the target user;

[0136] A target item for the target user is determined based on the probability value.

[0137] The probability value of the initial item relative to the target user can be understood as the probability of the target user clicking or purchasing the initial item.

[0138] In specific implementation, the target user characteristics of the target user and the target item characteristics of the initial item are input into the prediction model to obtain the probability value of the target user clicking or purchasing each initial item; subsequently, based on the probability value, the initial item with a higher probability value is selected as the target item of the target user.

[0139] In practical applications, the extraction of features in the polygonal graph, the embedding of the initial feature graph after concatenation, and the processing of features in isomorphic and heterogeneous graphs can all be implemented based on the neural network model. For example, the entire project recommendation method can be implemented based on the above recommendation model. The extraction of dense class features of edges in the polygonal graph and the extraction of initial project features in the initial project can be implemented through the feature extraction layer of the recommendation model. The embedding of the initial feature graph after concatenation can be implemented based on the embedding layer in the recommendation model. The isomorphic and heterogeneous graphs can also be implemented as the network layer in the recommendation model. The prediction model can also be implemented as the prediction output network layer of the recommendation model.

[0140] Compared with the feature transfer method, the project recommendation method provided in the embodiments of this specification designs an information transmission method between two different domains in the graph, combines the relationship information to transmit the information together, and uses the attention structure to realize a more fine-grained transmission between different nodes, so as to achieve adaptive information transmission for different nodes and different domains, thereby solving the problem of negative transfer; compared with the collaborative filtering method, this method introduces the capital relationship network and the social network to strengthen the features on the user side, and embeds the features to strengthen the information expression ability of the graph structure; thereby improving the recommendation effect of recommending target projects to target users.

[0141] In addition, when the project recommendation method is applied to cross-domain financial scenarios, a recommendation model under financial management and loan needs is adopted. Faced with the technical problem that user behavior is small and generally long-term and low-frequency conversion behavior makes it difficult to predict user needs (demand changes), this method introduces social networks (such as financial management scenarios) and capital networks (loans, etc.), and combines them with a multilateral graph structure. In this way, financial relationships and social relationships are introduced into the multilateral graph, and a domain-adaptive graph information transmission module is designed, so that the recommendation model can make accurate predictions on multiple domains, thereby improving the recommendation effect of recommending target projects to target users.

[0142] The project recommendation method provided in the embodiments of this specification captures information between users through a relationship feature graph that includes the resource relationships and social relationships of the target users, thereby better solving the problem of cold start; and can enhance the characteristics of the user side through the resource relationships and social relationships of the target users in the same domain or different domains, thereby greatly enhancing the user characteristics of the target users and avoiding the problem of being unable to make good project recommendations for new users by only considering feature sharing between domains.

[0143] Corresponding to the above method embodiment, this specification also provides a project recommendation device embodiment, Figure 3 FIG. 1 is a schematic diagram showing the structure of a project recommendation device provided by an embodiment of the present specification. Figure 3 As shown, the device comprises:

[0144] A target user determination module 302 is configured to determine a target user and initial items of different domains to be recommended to the target user;

[0145] The feature graph construction module 304 is configured to construct a relationship feature graph based on the target user, a first associated user having a resource relationship with the target user, and a second associated user having a social relationship with the target user, wherein the first associated user and / or the second associated user and the target user belong to the same domain or different domains;

[0146] A feature determination module 306 is configured to determine a target user feature of the target user and a target project feature of the initial project based on the relationship feature graph, the isomorphic graph, and the heterogeneous graph;

[0147] The project recommendation module 308 is configured to determine the target project of the target user from the initial projects based on the target user characteristics of the target user and the target project characteristics of the initial projects.

[0148] Optionally, the feature map construction module 304 is further configured to:

[0149] Based on the target user, determining a first associated user having a resource relationship with the target user and a second associated user having a social relationship with the target user;

[0150] constructing a first multi-edge graph and a second multi-edge graph based on the target user, the first associated user, the second associated user, the resource relationship between the target user and the first associated user, and the social relationship between the target user and the second associated user;

[0151] A relationship feature graph is constructed based on the first polygonal graph and the second polygonal graph.

[0152] Optionally, the feature map construction module 304 is further configured to:

[0153] Taking the target user and the first associated user as nodes and the resource relationship between the target user and the first associated user as edges, constructing a first multi-edge graph;

[0154] A second multi-edge graph is constructed by taking the target user and the second associated user as nodes and taking the social relationship between the target user and the second associated user as edges.

[0155] Optionally, the feature map construction module 304 is further configured to:

[0156] Extracting edge features of each edge in the first polygonal graph and edge features of each edge in the second polygonal graph;

[0157] Performing feature splicing on the edge feature of each edge in the first polygonal graph and the edge feature of each edge in the second polygonal graph to form an initial feature graph;

[0158] Based on a preset matrix, feature embedding is performed on the edge features of each edge in the initial feature graph to obtain a relationship feature graph.

[0159] Optionally, the feature determination module 306 is further configured to:

[0160] Determine, based on the relationship feature graph, edge features of the edge associated with the target user, user features of users in the same domain associated with the target user, and domain features of different domains associated with the target user;

[0161] The edge feature, the user feature, and the domain feature are transmitted through a homogeneous graph to obtain a fused feature of the target user;

[0162] The fusion features of the target user and the initial project features of the initial project are input into a heterogeneous graph to obtain the target user features of the target user and the target project features of the initial project.

[0163] Optionally, the feature determination module 306 is further configured to:

[0164] The edge feature, the user feature, and the domain feature are transmitted through an isomorphic graph to obtain a first feature, a second feature, and a third feature of the target user;

[0165] The first feature, the second feature, and the third feature are fused to obtain a fused feature of the target user.

[0166] Optionally, the feature determination module 306 is further configured to:

[0167] Transferring information of the edge feature through a homogeneous graph to obtain a first feature of the edge associated with the target user transferred to the target user;

[0168] The user feature is transmitted through the isomorphic graph to obtain a second feature transmitted to the target user by the user in the same domain associated with the target user;

[0169] The domain features are transmitted through the isomorphic graph to obtain a third feature transmitted from the different domains associated with the target user to the target user.

[0170] Optionally, the project recommendation module 308 is further configured to:

[0171] Inputting the target user characteristics of the target user and the target item characteristics of the initial item into a prediction model to obtain a probability value of the initial item relative to the target user;

[0172] A target item for the target user is determined based on the probability value.

[0173] Optionally, the device further comprises:

[0174] The model application module is configured as follows:

[0175] Inputting the target user and the initial item into a recommendation model to obtain a probability value of the initial item relative to the target user;

[0176] A target item for the target user is determined based on the probability value.

[0177] Optionally, the recommendation model is trained by the following steps:

[0178] determining a sample user, sample items of different domains recommended to the sample user, and a sample tag of the sample user relative to the sample item;

[0179] Building a sample relationship feature graph based on the sample user, a first sample user having a resource relationship with the sample user, and a second sample user having a social relationship with the sample user;

[0180] Determine the target sample user features of the sample user and the target sample item features of the sample item based on the sample relationship feature graph, the isomorphic graph, and the heterogeneous graph;

[0181] Inputting the target sample user features of the sample user and the target sample item features of the sample item into the prediction layer of the recommendation model to obtain the probability value of the sample item relative to the sample user;

[0182] A loss function of the recommendation model is calculated based on the probability value and the sample label, and the recommendation model is trained based on the loss function.

[0183] The project recommendation method provided in the embodiments of this specification captures information between users through a relationship feature graph that includes the resource relationships and social relationships of the target users, thereby better solving the problem of cold start; and can enhance the characteristics of the user side through the resource relationships and social relationships of the target users in the same domain or different domains, thereby greatly enhancing the user characteristics of the target users and avoiding the problem of being unable to make good project recommendations for new users by only considering feature sharing between domains.

[0184] The above is a schematic scheme of a project recommendation device of this embodiment. It should be noted that the technical scheme of the project recommendation device and the technical scheme of the project recommendation method described above are of the same concept, and the details not described in detail in the technical scheme of the project recommendation device can be found in the description of the technical scheme of the project recommendation method described above.

[0185] Figure 4The block diagram of a computing device 400 according to an embodiment of the present specification is shown. The components of the computing device 400 include but are not limited to a memory 410 and a processor 420. The processor 420 is connected to the memory 410 via a bus 430, and the database 450 is used to store data.

[0186] The computing device 400 also includes an access device 440 that enables the computing device 400 to communicate via one or more networks 460. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 440 may include one or more of any type of network interface (e.g., a network interface card (NIC)) wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a World Wide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, a near field communication (NFC) interface, and the like.

[0187] In one embodiment of the present specification, the above components of the computing device 400 and Figure 4 Other components not shown in the figure may also be connected to each other, for example, via a bus. It should be understood that Figure 4 The computing device structure block diagram shown is only for the purpose of illustration, and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0188] Computing device 400 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or PC. Computing device 400 may also be a mobile or stationary server.

[0189] The processor 420 is used to execute the following computer executable instructions, which, when executed by the processor, implement the steps of the above-mentioned project recommendation method.

[0190] The above is a schematic scheme of a computing device of this embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the project recommendation method described above are of the same concept, and the details not described in detail in the technical scheme of the computing device can be found in the description of the technical scheme of the project recommendation method described above.

[0191] An embodiment of the present specification also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the above-mentioned project recommendation method.

[0192] The above is a schematic scheme of a computer-readable storage medium of this embodiment. It should be noted that the technical scheme of the storage medium and the technical scheme of the project recommendation method described above are of the same concept, and the details not described in detail in the technical scheme of the storage medium can be found in the description of the technical scheme of the project recommendation method described above.

[0193] An embodiment of the present specification also provides a computer program, wherein when the computer program is executed in a computer, the computer is caused to execute the steps of the above-mentioned project recommendation method.

[0194] The above is a schematic scheme of a computer program of this embodiment. It should be noted that the technical scheme of the computer program and the technical scheme of the project recommendation method described above are of the same concept, and the details not described in detail in the technical scheme of the computer program can be found in the description of the technical scheme of the project recommendation method described above.

[0195] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0196] The computer instructions include computer program codes, which may be in source code form, object code form, executable files or some intermediate forms, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0197] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of this specification are not limited by the order of the actions described, because according to the embodiments of this specification, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of this specification.

[0198] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0199] The preferred embodiments of this specification disclosed above are only used to help explain this specification. The optional embodiments do not describe all the details in detail, nor do they limit the invention to only the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of the embodiments of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of this specification, so that technicians in the relevant technical field can well understand and use this specification. This specification is only limited by the claims and their full scope and equivalents.

Claims

1. A project recommendation method, include: Determine a target user and initial items of different domains to be recommended to the target user; Building a relationship feature graph based on the target user, a first associated user having a resource relationship with the target user, and a second associated user having a social relationship with the target user, wherein the first associated user and / or the second associated user and the target user belong to the same domain or different domains, wherein the resource relationship refers to a relationship of resource exchange with the target user, and the social relationship refers to a relationship of social exchange with the target user; Determine the target user feature of the target user and the target project feature of the initial project based on the relationship feature graph, the isomorphic graph, and the heterogeneous graph; Based on the target user characteristics of the target user and the target item characteristics of the initial items, the target item of the target user is determined from the initial items.

2. The project recommendation method according to claim 1, wherein the relationship feature graph is constructed based on the target user, the first associated user having a resource relationship with the target user, and the second associated user having a social relationship with the target user. include: Based on the target user, determining a first associated user having a resource relationship with the target user and a second associated user having a social relationship with the target user; constructing a first multi-edge graph and a second multi-edge graph based on the target user, the first associated user, the second associated user, the resource relationship between the target user and the first associated user, and the social relationship between the target user and the second associated user; A relationship feature graph is constructed based on the first polygonal graph and the second polygonal graph.

3. The project recommendation method according to claim 2, wherein the first multi-edge graph and the second multi-edge graph are constructed based on the target user, the first associated user, the second associated user, the resource relationship between the target user and the first associated user, and the social relationship between the target user and the second associated user. include: Taking the target user and the first associated user as nodes and the resource relationship between the target user and the first associated user as edges, constructing a first multi-edge graph; A second multi-edge graph is constructed by taking the target user and the second associated user as nodes and taking the social relationship between the target user and the second associated user as edges.

4. The project recommendation method according to claim 2, wherein the relationship feature graph is constructed based on the first polygon graph and the second polygon graph. include: Extracting edge features of each edge in the first polygonal graph and edge features of each edge in the second polygonal graph; Performing feature splicing on the edge feature of each edge in the first polygonal graph and the edge feature of each edge in the second polygonal graph to form an initial feature graph; Based on a preset matrix, feature embedding is performed on the edge features of each edge in the initial feature graph to obtain a relationship feature graph.

5. The project recommendation method according to claim 1, wherein the target user features of the target user and the target project features of the initial project are determined based on the relationship feature graph, the isomorphic graph and the heterogeneous graph. include: Determine, based on the relationship feature graph, edge features of the edge associated with the target user, user features of users in the same domain associated with the target user, and domain features of different domains associated with the target user; The edge feature, the user feature, and the domain feature are transmitted through a homogeneous graph to obtain a fused feature of the target user; The fusion features of the target user and the initial project features of the initial project are input into a heterogeneous graph to obtain the target user features of the target user and the target project features of the initial project.

6. The item recommendation method according to claim 5, wherein the edge features, the user features and the domain features are transmitted through an isomorphic graph to obtain a fusion feature of the target user, include: The edge feature, the user feature, and the domain feature are transmitted through an isomorphic graph to obtain a first feature, a second feature, and a third feature of the target user; The first feature, the second feature, and the third feature are fused to obtain a fused feature of the target user.

7. The item recommendation method according to claim 6, wherein the edge feature, the user feature and the domain feature are transmitted through an isomorphic graph to obtain the first feature, the second feature and the third feature of the target user, include: Transferring information of the edge feature through a homogeneous graph to obtain a first feature of the edge associated with the target user transferred to the target user; The user feature is transmitted through the isomorphic graph to obtain a second feature transmitted to the target user by the user in the same domain associated with the target user; The domain features are transmitted through the isomorphic graph to obtain a third feature transmitted from the different domains associated with the target user to the target user.

8. The project recommendation method according to claim 1, wherein the target project of the target user is determined from the initial projects based on the target user characteristics of the target user and the target project characteristics of the initial projects, include: Inputting the target user characteristics of the target user and the target item characteristics of the initial item into a prediction model to obtain a probability value of the initial item relative to the target user; A target item for the target user is determined based on the probability value.

9. The project recommendation method according to claim 1, after determining the target user and the initial projects of different domains to be recommended to the target user, further comprising: include: Inputting the target user and the initial item into a recommendation model to obtain a probability value of the initial item relative to the target user; A target item for the target user is determined based on the probability value.

10. The project recommendation method according to claim 9, wherein the recommendation model is trained by the following steps: determining a sample user, sample items of different domains recommended to the sample user, and a sample tag of the sample user relative to the sample item; Building a sample relationship feature graph based on the sample user, a first sample user having a resource relationship with the sample user, and a second sample user having a social relationship with the sample user; Determine the target sample user features of the sample user and the target sample item features of the sample item based on the sample relationship feature graph, the isomorphic graph, and the heterogeneous graph; Inputting the target sample user features of the sample user and the target sample item features of the sample item into the prediction layer of the recommendation model to obtain the probability value of the sample item relative to the sample user; A loss function of the recommendation model is calculated based on the probability value and the sample label, and the recommendation model is trained based on the loss function.

11. A project recommendation device, include: a target user determination module, configured to determine a target user and initial items of different domains to be recommended to the target user; A feature graph construction module is configured to construct a relationship feature graph based on the target user, a first associated user having a resource relationship with the target user, and a second associated user having a social relationship with the target user, wherein the first associated user and / or the second associated user belongs to the same domain or different domains as the target user, wherein the resource relationship refers to a relationship of resource exchange with the target user, and the social relationship refers to a relationship of social exchange with the target user; A feature determination module, configured to determine a target user feature of the target user and a target project feature of the initial project based on the relationship feature graph, the isomorphic graph, and the heterogeneous graph; The project recommendation module is configured to determine a target project of the target user from the initial projects based on a target user feature of the target user and a target project feature of the initial project.

12. A computing device, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the project recommendation method described in any one of claims 1 to 10 are implemented.

13. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the project recommendation method according to any one of claims 1 to 10.

Images