Double-channel perception multi-behavior recommendation method, system and device and medium

Abstract

The invention discloses a dual-channel perception multi-behavior recommendation method, system and device and a medium, and the method comprises the steps: obtaining a multi-behavior heterogeneous graph, inputting the multi-behavior heterogeneous graph into a recommendation model, carrying out the node embedding processing of the multi-behavior heterogeneous graph through a node-level multi-head attention network module, and obtaining a behavior node embedding set; performing residual connection processing on the behavior node embedding set through a behavior level graph attention layer module fusing residual connection to obtain a first node embedding set; edge embedding display adding processing is carried out on the multi-behavior heterogeneous graph through the fusion display edge embedding graph attention network, and a second node embedding set is obtained; performing joint prediction processing on the first node embedding set and the second node embedding set through a joint prediction module, and outputting to obtain a prediction score; and performing recommendation processing according to the prediction score to obtain a multi-behavior recommendation result. The embodiment of the invention can improve the accuracy of multi-behavior recommendation, and can be widely applied to the technical field of machine learning.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to a dual-channel perception multi-behavior recommendation method, system, device and medium. Background Art

[0002] The rapid development of information technology has not only improved people's living standards, but also led to an explosive growth of network information. As an information filtering method, the recommendation system can provide users with personalized content and improve user experience. In related technologies, the recommendation method uses the prediction and training paradigm for recommendation, but ignores the high-order structural information in the data, which limits the system's ability to effectively obtain complex neighbor relationships, resulting in low efficiency and accuracy of recommendations.

[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the invention

[0004] The main purpose of the embodiments of the present application is to propose a dual-channel perception multi-behavior recommendation method, system, device and medium, which can improve the accuracy of recommendations.

[0005] To achieve the above purpose, an embodiment of the present application proposes a dual-channel perception multi-behavior recommendation method, the method comprising:

[0006] Acquire a multi-behavior heterogeneous graph, and input the multi-behavior heterogeneous graph into a recommendation model, wherein the recommendation model includes a behavior-layer graph attention network, a graph attention network fused with explicit edge embedding, and a joint prediction module, and the behavior-layer graph attention network includes a node-level multi-head attention network module and a behavior-level graph attention layer module fused with residual connections;

[0007] Performing node embedding processing on the multi-behavior heterogeneous graph through the node-level multi-head attention network module to obtain a behavior node embedding set;

[0008] Performing residual connection processing on the behavior node embedding set through the behavior-level graph attention layer module fused with residual connection to obtain a first node embedding set;

[0009] Performing edge embedding display adding processing on the multi-behavior heterogeneous graph through the graph attention network of the fused display edge embedding to obtain a second node embedding set;

[0010] Performing joint prediction processing on the first node embedding set and the second node embedding set by the joint prediction module, and outputting a prediction score;

[0011] Recommendation processing is performed according to the prediction scores to obtain multiple behavior recommendation results.

[0012] In some embodiments, performing node embedding processing on the multi-behavior heterogeneous graph by the node-level multi-head attention network module to obtain a behavior node embedding set includes the following steps:

[0013] Performing high-dimensional feature mapping and attention calculation processing on the multi-behavior heterogeneous graph to obtain initial node embedding and initial neighbor node weights;

[0014] Performing nonlinear transformation and normalization processing on the initial neighbor node weights to obtain target neighbor node weights;

[0015] The initial node embedding is updated and multi-head attention calculation is performed according to the target neighbor node weight to obtain the behavior node embedding set.

[0016] In some embodiments, the residual connection processing is performed on the behavior node embedding set by the behavior-level graph attention layer module through the fused residual connection to obtain a first node embedding set, including the following steps:

[0017] Performing multi-layer perception and nonlinear transformation processing on the behavior node embedding set to obtain initial behavior node weights;

[0018] Normalizing the initial behavior node weight to obtain a target behavior node weight;

[0019] Aggregating the behavior node embedding set according to the target behavior node weight to obtain an aggregated embedding set;

[0020] Perform residual calculation processing on the aggregated embedding set according to the residual parameter to obtain the first node embedding set.

[0021] In some embodiments, the step of performing edge embedding display adding processing on the multi-behavior heterogeneous graph through the graph attention network of the fused display edge embedding to obtain a second node embedding set includes the following steps:

[0022] Performing behavior embedding and graph attention calculation processing on the multi-behavior heterogeneous graph to obtain an initial attention score;

[0023] Performing nonlinear transformation and normalization processing on the initial attention score to obtain a target attention score;

[0024] Update the node embeddings of the multi-behavior heterogeneous graph according to the target attention scores to obtain an edge embedding set;

[0025] Perform multi-head attention calculation and residual connection processing on the edge embedding set to obtain the second node embedding set.

[0026] In some embodiments, performing joint prediction processing on the first node embedding set and the second node embedding set by the joint prediction module and outputting a prediction score comprises the following steps:

[0027] Performing inner product calculation processing of user node embedding and item node embedding on the first node embedding set to obtain a first score;

[0028] Performing inner product calculation processing of the user node embedding and the item node embedding on the second node embedding set to obtain a second score;

[0029] The first score and the second score are combined according to the learnable parameter to obtain the predicted score.

[0030] In some embodiments, before inputting the multi-behavior heterogeneous graph into the recommendation model, the method further includes pre-training the recommendation model, specifically including the following steps:

[0031] Obtain positive sample data and negative sample data;

[0032] Inputting the positive sample data and the negative sample data into the recommendation model for interactive prediction respectively, to obtain a positive sample prediction score and a negative sample prediction score;

[0033] Performing loss processing on the positive sample prediction score and the negative sample prediction score by using a loss function to obtain a loss value;

[0034] The parameters of the recommendation model are updated according to the loss value to obtain the trained recommendation model.

[0035] In some embodiments, obtaining a heterogeneous graph of multiple behaviors includes the following steps:

[0036] Get user-project interaction records;

[0037] Behavior analysis and graph construction processing are performed on the user-project interaction records to obtain the multi-behavior heterogeneous graph.

[0038] To achieve the above purpose, another aspect of the embodiment of the present application proposes a dual-channel perception multi-behavior recommendation system, the system comprising:

[0039] The first module is used to obtain a multi-behavior heterogeneous graph and input the multi-behavior heterogeneous graph into a recommendation model, wherein the recommendation model includes a node-level multi-head attention network module, a behavior-level graph attention layer fused with residual connections, a graph attention network module fused with explicit edge embedding, and a joint prediction module;

[0040] The second module is used to perform node embedding processing on the multi-behavior heterogeneous graph through the node-level multi-head attention network module to obtain a behavior node embedding set;

[0041] A third module is used to perform residual connection processing on the behavior node embedding set through the behavior-level graph attention layer fused with residual connection to obtain a first node embedding set;

[0042] A fourth module is used to perform edge embedding display adding processing on the multi-behavior heterogeneous graph through the graph attention network of the fused display edge embedding to obtain a second node embedding set;

[0043] A fifth module, configured to perform joint prediction processing on the first node embedding set and the second node embedding set through the joint prediction module, and output a prediction score;

[0044] The sixth module is used to perform recommendation processing according to the prediction score to obtain multiple behavior recommendation results.

[0045] To achieve the above-mentioned purpose, another aspect of an embodiment of the present application provides an electronic device, the electronic device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the above-mentioned method when executing the computer program.

[0046] To achieve the above objective, another aspect of an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method described above is implemented.

[0047] The embodiments of the present application include at least the following beneficial effects: The present application provides a dual-channel-aware multi-behavior recommendation method, system, device and medium. The scheme obtains a multi-behavior heterogeneous graph, inputs the multi-behavior heterogeneous graph into the recommendation model, performs node embedding processing on the multi-behavior heterogeneous graph through a node-level multi-head attention network module, and obtains a behavior node embedding set; performs residual connection processing on the behavior node embedding set through a behavior-level graph attention layer module that fuses residual connections to obtain a first node embedding set; performs edge embedding display addition processing on the multi-behavior heterogeneous graph through a graph attention network that fuses display edge embedding to obtain a second node embedding set; performs joint prediction processing on the first node embedding set and the second node embedding set through a joint prediction module, and outputs a prediction score; performs recommendation processing based on the prediction score to obtain a multi-behavior recommendation result. The embodiments of the present application use a graph attention network to aggregate neighbor node information, and at the same time adopt a dual-channel strategy at the behavior level, using a behavior layer graph attention network and a graph attention network that fuses explicit edge embedding to obtain behavior features, which enables the recommendation model to obtain subtle differences in behavior features, significantly enhances the expressiveness of the model, and thus improves the efficiency and accuracy of recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a flow chart of a dual-channel perception multi-behavior recommendation method provided by an embodiment of the present application;

[0049] Figure 2 It is a schematic diagram of the overall framework of a recommendation model provided in an embodiment of the present application;

[0050] Figure 3 It is a schematic diagram of a framework of a node-level attention layer provided in an embodiment of the present application;

[0051] Figure 4 It is a schematic diagram of a behavior-level attention layer framework provided in an embodiment of the present application;

[0052] Figure 5 It is a schematic diagram of a graph attention network module framework integrating explicit edge embedding provided in an embodiment of the present application;

[0053] Figure 6 It is a structural diagram of a dual-channel perception multi-behavior recommendation system provided in an embodiment of the present application;

[0054] Figure 7 It is a schematic diagram of the hardware structure of the electronic device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in conjunction with the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of the present application. They are only examples of systems and methods consistent with some aspects of the embodiments of the present application as detailed in the attached claims.

[0056] It is understood that the terms "first", "second", etc. used in this application can be used to describe various concepts in this article, but unless otherwise specified, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another concept. For example, without departing from the scope of the embodiment of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the words "if" and "if" as used herein can be interpreted as "at the time of" or "when" or "in response to determination".

[0057] The terms "at least one", "multiple", "each", "any", etc. used in this application, at least one includes one, two or more, multiple includes two or more, each refers to each of the corresponding multiple, and any refers to any one of the multiple.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0059] The rapid development of information technology has not only improved people's living standards, but also led to an explosive growth of network information. Therefore, solving the problem of information overload and extracting effective information from large amounts of data based on user interests has become a research focus for enterprises and organizations. As an information filtering method, the recommendation system can provide users with personalized content and improve user experience. However, traditional recommendation methods use prediction and training paradigms that ignore high-order structural information in the data. This limits the system's ability to effectively obtain complex neighbor relationships, thereby ignoring items that do not directly interact with users but meet their potential needs.

[0060] In the related technology, there is a recommendation method based on graph neural network, that is, based on graph neural network, each node can obtain high-order neighbor information to process structured data and structural information. However, recommendation systems often encounter data sparsity or even cold start problems, especially in practical applications that rely heavily on historical user behavior data. In practical applications, it is found that the existing multi-behavior recommendation methods fail to utilize or fully utilize rich auxiliary information to improve recommendation performance, and the node feature learning ability of traditional graph neural networks is weak. The processing method of multi-behavior information is simple, which hinders the effective use of auxiliary behavior information and affects the efficiency and accuracy of recommendations.

[0061] In view of this, a dual-channel-aware multi-behavior recommendation method, system, device and medium are provided in the embodiments of the present application. The scheme obtains a multi-behavior heterogeneous graph, inputs the multi-behavior heterogeneous graph into a recommendation model, performs node embedding processing on the multi-behavior heterogeneous graph through a node-level multi-head attention network module to obtain a behavior node embedding set; performs residual connection processing on the behavior node embedding set through a behavior-level graph attention layer module that integrates residual connections to obtain a first node embedding set; performs edge embedding display addition processing on the multi-behavior heterogeneous graph through a graph attention network that integrates explicit edge embedding to obtain a second node embedding set; performs joint prediction processing on the first node embedding set and the second node embedding set through a joint prediction module, and outputs a prediction score; performs recommendation processing based on the prediction score to obtain a multi-behavior recommendation result. The embodiments of the present application adopt a graph attention network to aggregate neighbor node information, and at the same time adopt a dual-channel strategy at the behavior level, using a behavior layer graph attention network and a graph attention network that integrates explicit edge embedding to obtain behavior features, which enables the recommendation model to obtain subtle differences in features between behaviors, significantly enhancing the expressiveness of the model, thereby improving the efficiency and accuracy of recommendations.

[0062] The dual-channel-aware multi-behavior recommendation method provided in the embodiment of the present application relates to the field of machine learning technology. The dual-channel-aware multi-behavior recommendation method provided in the embodiment of the present application can be applied to a terminal, can also be applied to a server, and can also be software running in a terminal or a server. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, and a car terminal, etc., but is not limited to this; the server side can be configured as an independent physical server, or it can be configured as a server cluster or a distributed system composed of multiple physical servers, and can also be configured to provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms and other basic cloud computing services. The server can also be a node server in a blockchain network; the software can be an application of a multi-behavior recommendation method that implements dual-channel perception, etc., but is not limited to the above forms.

[0063] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0064] It should be noted that in each specific implementation of the present application, when it comes to the need to perform relevant processing based on data related to user identity or characteristics such as user information, user behavior data, user historical data, and user location information, the user's permission or consent will be obtained first, and the collection, use, and processing of these data will comply with relevant laws, regulations, and standards. In addition, when the embodiment of the present application needs to obtain the user's sensitive personal information, the user's separate permission or consent will be obtained through a pop-up window or by jumping to a confirmation page. After clearly obtaining the user's separate permission or consent, the necessary user-related data for the normal operation of the embodiment of the present application will be obtained.

[0065] Figure 1 is an optional flowchart of a dual-channel perception multi-behavior recommendation method provided in an embodiment of the present application. Figure 1 The method may include but is not limited to steps S101 to S106.

[0066] Step S101, obtaining a multi-behavior heterogeneous graph, and inputting the multi-behavior heterogeneous graph into a recommendation model;

[0067] Step S102, performing node embedding processing on the multi-behavior heterogeneous graph through the node-level multi-head attention network module to obtain a behavior node embedding set;

[0068] Step S103, performing residual connection processing on the behavior node embedding set through the behavior-level graph attention layer module fused with residual connection to obtain a first node embedding set;

[0069] Step S104, performing edge embedding display adding processing on the multi-behavior heterogeneous graph through the graph attention network of the fused display edge embedding to obtain a second node embedding set;

[0070] Step S105, performing joint prediction processing on the first node embedding set and the second node embedding set through the joint prediction module, and outputting a prediction score;

[0071] Step S106, performing recommendation processing according to the prediction score to obtain multiple behavior recommendation results.

[0072] In the steps S101 to S106 shown in the embodiment of the present application, the multi-behavior heterogeneous graph is input into the recommendation model, and the scores of the multi-behavior heterogeneous graph are predicted respectively through the dual channels of the recommendation model, namely the behavior layer graph attention network and the graph attention network fused with explicit edge embedding, and finally the final prediction score is output through the joint prediction module, so as to recommend items to users according to the prediction scores and obtain multi-behavior recommendation results. Please refer to Figure 2 , Figure 2 A schematic diagram of the overall framework of a recommendation model provided in an embodiment of the present application, wherein a multi-behavior heterogeneous graph may include a user-item interaction graph under multiple behaviors, in which j represents an item, N represents an item set, and V represents an item group. j1 、V j2 and V j3 Represent three different projects, t 1 ,t 2 and t m They represent different behaviors respectively, m is the number of behaviors, e tm Represents the relationship between users and items under different behaviors. By inputting multiple heterogeneous graphs into the recommendation module, the recommendation model includes a behavior-level graph attention network, a graph attention network fused with explicit edge embedding, and a joint prediction module. The behavior-level graph attention network includes a node-level multi-head attention network module and a behavior-level graph attention layer module fused with residual connections. Specifically, refer to Figure 2 In the recommendation model, the behavior layer graph attention network can include basic modules such as linear transformation layer, multi-level graph attention model, residual module, etc. The graph attention network fused with explicit edge embedding can include basic modules such as graph attention module fused with variable embedding and residual module, among which, O 1 , Z 1 They represent the output of each layer of the graph attention network, and the score Y is obtained by outputting the behavior layer graph attention network and the graph attention network fused with explicit edge embedding. 2, and the output scores are jointly predicted to obtain the final prediction score Y. The predicted scores can be used to sort the items and recommend them to users based on the sorting results. The embodiment of the present application adopts a dual-channel strategy for the graph attention network, which can simultaneously obtain the impact of behaviors on nodes by using a multi-level graph attention network module that integrates residual connections and a graph attention network module that integrates explicit edge embeddings, and also uses a joint prediction module to effectively integrate these components together, thereby improving the accuracy and efficiency of predictions.

[0073] In step S101 of some embodiments, obtaining the multiple lines as a heterogeneous graph includes the following steps:

[0074] Get user-project interaction records;

[0075] Behavior analysis and graph construction processing are performed on the user-project interaction records to obtain the multi-behavior heterogeneous graph.

[0076] In an embodiment of the present application, a user-project interaction record is obtained, and the interaction record is whether the user interacts with the project. If an interaction occurs, the record is 1, otherwise the record is 0. The embodiment of the present application performs behavioral analysis on the user-project interaction record to analyze the specific behaviors of the user's interaction with the project, such as shopping, adding to a shopping cart, clicking, etc., so that different user-project interaction bipartite graphs can be created according to the interaction records under different behaviors. This type of graph can be reflected in an adjacency matrix in the specific implementation program. Then, a multi-behavior heterogeneous graph is obtained based on the user-project interaction bipartite graph under multiple different behaviors. Multiple behaviors can be naturally reflected as different types of edges on the heterogeneous graph. Suppose the multi-behavior heterogeneous graph data is G=(U,V), where U represents the user node set u∈U; V represents the project node set v∈V. The interaction between u and v forms an edge e t (u,v)∈E, where t∈M represents the behavior type to which the edge belongs, M is the set of all behavior types, and E represents the set of all edges. The embodiment of the present application performs behavioral analysis on the user-project interaction records so that the constructed multi-behavior heterogeneous graph has various types of edge information, which can improve the performance of the model, and constructs a heterogeneous graph using rich auxiliary information in actual application scenarios, thereby enhancing the learning ability of the model. In addition, compared with the model that only uses a single behavior, the embodiment of the present application uses auxiliary behaviors to enrich the data volume of the model, thereby better handling the problem of data sparsity.

[0077] In step S102 of some embodiments, performing node embedding processing on the multi-behavior heterogeneous graph by the node-level multi-head attention network module to obtain a behavior node embedding set includes the following steps:

[0078] Performing high-dimensional feature mapping and attention calculation processing on the multi-behavior heterogeneous graph to obtain initial node embedding and initial neighbor node weights;

[0079] Performing nonlinear transformation and normalization processing on the initial neighbor node weights to obtain target neighbor node weights;

[0080] The initial node embedding is updated and multi-head attention calculation is performed according to the target neighbor node weight to obtain the behavior node embedding set.

[0081] In the embodiment of the present application, high-dimensional feature mapping and attention calculation processing are performed on the multi-behavior heterogeneous graph. Specifically, the importance of each neighbor node is obtained for the multi-behavior heterogeneous graph through the graph attention network, thereby enhancing the learning ability of node embedding, and obtaining the initial node embedding and initial neighbor node weight. The calculation formula of the initial neighbor node weight is as follows:

[0082]

[0083] In the formula, represents the initial neighbor node weight, t represents the behavior type, i represents the target node, j represents the neighbor node, and h i and h j Represents the initial node embedding. Considering the heterogeneity of heterogeneous graph nodes, different nodes may be located in different feature spaces, so the feature transformation matrix W is used to unify different types of nodes in the same feature space. || represents the connection operation. a represents the attention calculation operation that maps high-dimensional features to weighted real numbers. Among them, the initial neighbor node weight represents the initial weight of the neighbor node and the target node. After calculating the initial neighbor node weight, the weight needs to be nonlinearly transformed and normalized to obtain the target neighbor node weight. The calculation formula of the target neighbor node weight is as follows:

[0084]

[0085] In the formula, represents the target neighbor node weight, σ is the LeakyReLU activation function, N t(i) represents the set of nodes that interact with i under behavior t, that is, the neighbor nodes of i under behavior t. Finally, the initial node embedding is updated and multi-head attention is calculated according to the weight of the target neighbor node. The expression for updating the node embedding is as follows:

[0086]

[0087] In the formula, Represents the updated node embedding. Considering the scale-free and large variance characteristics of heterogeneous graphs, node-level attention can be extended to multi-head attention to enhance the stability of training. Specifically, the multi-head attention calculation formula for the number of attention heads K is as follows:

[0088]

[0089] In the formula, represents the behavior node embedding set. For example, given the behavior set {θ1, θ2…θk}, we can get the behavior node embedding set {H θ1 , H θ2 …H θk}. See Figure 3 In the data processing stage, the embodiments of the present application can initialize the nodes on the bipartite graphs of user-item interactions under each behavior in the multi-behavior heterogeneous graph. It is worth mentioning that the embedding of the input and output of each layer of the model is the same for a single node, rather than the node having an embedding under each behavior. Figure 3 K represents the number of layers in the model, t represents the behavior type, and t 1 Represents the first behavior, for the interaction diagram U i1 、V j1 、V j2 、V j3 They represent nodes, h i1 、h j1 、h j2 、h j3 They represent the embedding corresponding to the node, a i1j1 、a i1j2 、a i1j3Respectively represent the attention calculation parameters. The initial embedding of the node can come from the random initialization of the framework, the pre-training of a simple model, the pre-training of a language model, etc. First, the parameters are initialized. The trainable parameters in this module can be randomly initialized by the framework, such as the feature transformation matrix W and the attention calculation parameter a that maps high-dimensional features to weighted real numbers, etc. The trainable parameters of other modules are the same as above. In the embodiment of the present application, under behavior t, a graph attention calculation operation is performed on the target node i. Specifically, for i and a neighbor j of i, different feature transformation matrices are used to unify the feature space, and then tensors are spliced. The spliced ​​tensor is calculated with the attention parameters, and finally the initial neighbor node weight is obtained. The initial neighbor node weight represents the weight of the neighbor node j to the target node i under behavior t. Then, the obtained initial neighbor node weight is used to perform a softmax calculation on the weight values ​​of the target node i and its neighbors to obtain the target neighbor node weight under behavior t. Finally, multiply the target neighbor node weight by the initial embedding of the neighbor node, and then add up all the new embeddings obtained in this way to get the updated node embedding. In order to prevent training instability, a multi-head attention mechanism is introduced, and the above steps are repeated k times to obtain k embedding expressions hi of the target node i. Then, these k hi are fused, which can be tensor connection, average calculation, or more complex aggregation methods, to output the embedding of the target node i under a specific behavior t, and obtain the behavior node embedding set for the embeddings under multiple behaviors. The embodiment of the present application can obtain the behavior node embedding set of the target node under a specific behavior through the node-level multi-head attention network module, which provides a computational basis for subsequent score prediction.

[0090] In step S103 of some embodiments, the residual connection processing is performed on the behavior node embedding set by the behavior-level graph attention layer module fused with residual connection to obtain a first node embedding set, including the following steps:

[0091] Performing multi-layer perception and nonlinear transformation processing on the behavior node embedding set to obtain initial behavior node weights;

[0092] Normalizing the initial behavior node weight to obtain a target behavior node weight;

[0093] Aggregating the behavior node embedding set according to the target behavior node weight to obtain an aggregated embedding set;

[0094] Perform residual calculation processing on the aggregated embedding set according to the residual parameter to obtain the first node embedding set.

[0095] In the embodiment of the present application, a graph attention network is used to calculate the importance of different behaviors to nodes by mapping behaviors to different types of edges on a heterogeneous graph, and the residual connection module is combined to retain the information of the node itself. Specifically, the initial behavior node weight is obtained by performing multi-layer perception and nonlinear transformation processing on the behavior node embedding set. The calculation formula of the initial behavior node weight is as follows:

[0096]

[0097] In the formula, represents the initial behavior node weight, W t 、b t is the linear change parameter corresponding to behavior t. tanh is the activation function, and q is an attention parameter matrix that maps high-dimensional features to weighted real numbers. After obtaining the importance of different behaviors to nodes, the initial behavior node weights are normalized to obtain the final weight of the behavior to the node, that is, the target behavior node weight. The calculation formula for the target behavior node weight is as follows:

[0098]

[0099] In the formula, represents the target behavior node weight, and M represents the set of all behavior types. Then, according to the weight values ​​of different behaviors, the behavior node embedding set of the node-level graph attention network layer is aggregated. The aggregation formula is as follows:

[0100]

[0101] In the formula, Z i Represents aggregate embedding. Finally, the residual connection module is used to prevent the over-smoothing caused by the graph neural network from causing the loss of node information. The calculation formula for the first node embedding set is as follows:

[0102]

[0103] In the formula, represents the behavioral attention output of the l-th layer model. For the l+1-th layer model, the target node i is embedded as The final representations are combined to get the first node embedding set, and res is the residual parameter. Figure 4, after the previous module, the embodiment of the present application obtains the embedding of each node under a specific behavior t. There are m behaviors, there are m bipartite graphs, and the nodes also have m embeddings to obtain a behavior node embedding set as the input of this module. Specifically, the behavior node embedding of node i under a specific behavior t is passed through a multi-layer perceptron (MLP), and then a nonlinear change is performed, and finally multiplied with an attention parameter matrix q to obtain the initial behavior node weight, which represents the importance of behavior t to node i. Then the initial behavior node weight obtained under each behavior of node i is softmax calculated, and the target behavior node weight obtained after softmax is multiplied by the behavior node embedding under the corresponding behavior, and finally added to obtain the final embedding of i, that is, the aggregate embedding. For the model of the l+1th layer, the embodiment of the present application has the node embedding ol of the lth layer input, and the output zl+1 of this layer. Multiply o with the residual parameter res, and then add it to zl+1 to obtain the final input ol+1 of this module at this layer, and finally obtain the first node embedding set, which is the embedding set obtained by processing the behavior layer graph attention network. Among them, the residual parameter is a hyperparameter between 0-1. The embodiment of the present application calculates importance through a behavior layer graph attention network, and can explore the impact of behaviors on interactions within heterogeneous graphs through a dynamic method, thereby improving the efficiency of multi-behavior recommendations.

[0104] In step S104 of some embodiments, performing edge embedding display adding processing on the multi-behavior heterogeneous graph through the graph attention network of the fused display edge embedding to obtain a second node embedding set includes the following steps:

[0105] Performing behavior embedding and graph attention calculation processing on the multi-behavior heterogeneous graph to obtain an initial attention score;

[0106] Performing nonlinear transformation and normalization processing on the initial attention score to obtain a target attention score;

[0107] Update the node embeddings of the multi-behavior heterogeneous graph according to the target attention scores to obtain an edge embedding set;

[0108] Perform multi-head attention calculation and residual connection processing on the edge embedding set to obtain the second node embedding set.

[0109] In the embodiment of the present application, the behavior embedding and graph attention calculation processing are performed on the multi-behavior heterogeneous graph by fusing the graph attention network with edge embedding, and the attention calculation between nodes can be performed by explicitly adding edge embedding, and the static type information is integrated into the node to obtain the initial attention score. The calculation formula of the initial attention score is as follows:

[0110]

[0111] In the formula, represents the initial attention score, W b is the unified weight matrix for edge embedding in the unified vector feature space, r t is the embedding of behavior type t. After calculating the initial attention score, it needs to be nonlinearly transformed and normalized to obtain the target attention score. The calculation formula of the target attention score is as follows:

[0112]

[0113] In the formula, Represents the target attention score. Then, the node embedding of the multi-behavior heterogeneous graph is updated according to the target attention score to obtain the edge embedding set, where the edge embedding calculation formula is as follows:

[0114]

[0115] In the formula, represents edge embedding, h j Represents the node embedding of a heterogeneous graph with multiple behaviors. Finally, similar to the node-level graph attention network module, the edge embedding set is processed by a multi-head attention mechanism and residual connection to obtain a second node embedding set. The second node embedding set is the embedding set obtained by fusing the graph attention network processing with the edge embedding. See Figure 5 For each behavior t, the embodiment of the present application initializes a vector as the embedding of t, representing the characteristics of behavior t, which can be obtained by random initialization of the framework. Figure 5 In the above code, t indicates the behavior type. m represents the mth behavior, m represents the total number of behaviors, for the interaction graph U i1 、V j1 、V j2 、V j3 They represent nodes, h i1 、h j1 、h j2 、h j3 They represent the embedding corresponding to the node, a i1j1 、a i1j2 、a i1j3 They represent the attention calculation parameters, e tm Indicates that in behavior mEdge embedding under. Perform graph attention calculation operation for target node i. Specifically, for i and a neighbor j of i, as well as behavior embedding, different feature matrices W are used to process the unified feature space, and then tensor splicing is performed. The spliced ​​tensor is calculated with the attention parameter a, and finally the initial attention score is obtained, which indicates the importance of neighbor j and behavior t to i when behavior t occurs. Then, under specific behavior t, the initial attention scores of all neighbors j are softmax calculated to obtain a new weight, namely the target attention score. Multiply the node embedding of neighbor node j by the target attention score, and then add up all the new embeddings obtained in this way to obtain the second node embedding of target node i, thereby obtaining a second node embedding set. The embodiment of the present application characterizes the characteristics of the behavior itself by integrating the graph attention network that displays edge embeddings, thereby enhancing the expressive power of the model and improving the efficiency of recommendation.

[0116] In step S105 of some embodiments, performing joint prediction processing on the first node embedding set and the second node embedding set by the joint prediction module and outputting a prediction score comprises the following steps:

[0117] Performing inner product calculation processing of user node embedding and item node embedding on the first node embedding set to obtain a first score;

[0118] Performing inner product calculation processing of the user node embedding and the item node embedding on the second node embedding set to obtain a second score;

[0119] The first score and the second score are combined according to the learnable parameter to obtain the predicted score.

[0120] In the embodiment of the present application, after being processed by the behavior layer graph attention network and the graph attention network fused with explicit edge embedding, two sets of node embeddings are finally obtained, namely the first node embedding set and the second node embedding set. The embodiment of the present application uses a joint prediction module to comprehensively utilize the first node embedding set and the second node embedding set. and denote the final output node embeddings of users and items from the multi-level graph attention module, respectively. Similarly, let and Represents the final output node embeddings of users and items from the graph attention network module fused with explicit edge embeddings. The prediction scores of these two networks can be expressed as the inner product, where the first score and the second score are calculated as follows:

[0121]

[0122]

[0123] In the formula, y 1 represents the first fraction, y 2 Represents the second score. Then a learnable parameter λ is used to combine the two scores to get the final prediction score y, which is calculated as follows:

[0124] y=λy 1 +(1-λ)y 2 ;

[0125] The embodiment of the present application integrates node neighbor information through a joint prediction module, and integrates behavioral semantics from both static and dynamic aspects, which can improve the learning ability of the model and thus improve the efficiency and accuracy of multi-behavior recommendations.

[0126] In some embodiments, before inputting the multi-behavior heterogeneous graph into the recommendation model, the method further includes pre-training the recommendation model, specifically including the following steps:

[0127] Obtain positive sample data and negative sample data;

[0128] Inputting the positive sample data and the negative sample data into the recommendation model for interactive prediction respectively, to obtain a positive sample prediction score and a negative sample prediction score;

[0129] Performing loss processing on the positive sample prediction score and the negative sample prediction score by using a loss function to obtain a loss value;

[0130] The parameters of the recommendation model are updated according to the loss value to obtain the trained recommendation model.

[0131] In the embodiment of the present application, the recommendation model is trained, and Bayesian personalized ranking (bpr loss) is used as the loss function. It can use positive samples and negative samples at the same time, so that the model is more inclined to predict items that generate interactions, but also more inclined to avoid predicting items that do not interact. The expression of the loss function of the embodiment of the present application is as follows:

[0132] Loss=-log(σ(y(u,i)-y(u,j)))+η||L2|| 2 ;

[0133] Wherein, y(u,i) is the predicted score of positive sample i, and y(u,j) is the predicted score of negative sample j. σ is the sigmoid function. L2 is the L2 regularization value, and η is the regularization parameter to prevent overfitting. It should be noted that in order to alleviate the common overfitting problem in deep learning models, the embodiment of the present application adds a dropout module to the node aggregation and message passing parts during the training process. In the node aggregation part, dropout randomly deletes certain interactive nodes, and in the message passing module, it randomly discards some information during the information aggregation process. The embodiment of the present application can improve the efficiency of model training through loss function and random dropout processing.

[0134] The following is a detailed description of the embodiments of the present invention with reference to specific application examples:

[0135] The embodiment of the present application can be applied to recommendation application scenarios, such as music recommendation, product recommendation and other application scenarios. The recommendation model of the dual-channel perceived multi-behavior recommendation method of the embodiment of the present application is compared with other benchmark models under different data sets, and the recall rate (Recall) and NDCG are used as the evaluation indicators of the experiment. Other benchmark models include RGCN model, NGCF-MB model, etc. The embodiment of the present application is implemented using Pytorch, and the Adam optimizer is set and the training batch size is set to 1024. For the negative samples used by the bpr loss function, a random sampling strategy is used to collect negative samples. For the learning rate and L2 regularization coefficient, a grid search strategy is used, and they are adjusted in [1e-4, 3e-4, 1e-5, 3e-5, 1e-6] and [1e-3, 1e-4, 1e-5, 1e-6, 1e-7] respectively. In order to prevent overfitting, the embodiment of the present application sets an early stopping mechanism, and the training will stop when the index on the validation set does not increase by more than 50 times. And the number of attention heads is set to 6, and the model is stacked into two layers. By comparing the recommendation model of the embodiment of the present application with other graph neural network-based models that can extract information from multi-behavior data, the recommendation model of the embodiment of the present application performs best, which can be explained in three ways. First, the baseline method only focuses on the implicit features of the behavior type during the information transfer between nodes, while ignoring the static features of the behavior type. These static features belong to the inherent properties of the type, and these properties should remain consistent regardless of the changes in the connected nodes. Secondly, the recommendation model of the embodiment of the present application adopts a dual-channel strategy to obtain behavioral features, and its characterization ability is better than other multi-behavior recommendation models. Third, the recommendation model of the embodiment of the present application uses residual connections to avoid the common over-smoothing phenomenon in graph neural networks, so that nodes can better retain their own characteristics. It can be seen that the dual-channel-aware multi-behavior recommendation method of the embodiment of the present application has made progress in some typical indicators of the recommendation system. By adopting a dual-channel strategy of the graph attention network, a multi-level graph attention network module with fused residual connections and a graph attention network module with fused explicit edge embedding are used to simultaneously obtain the impact of behavior on nodes. In addition, the embodiment of the present application also uses a joint prediction module to effectively integrate these components together.

[0136] See also Figure 6 The embodiment of the present application also provides a dual-channel perception multi-behavior recommendation system, which can implement the above-mentioned dual-channel perception multi-behavior recommendation method, and the system includes:

[0137] The first module 601 is used to obtain a multi-behavior heterogeneous graph and input the multi-behavior heterogeneous graph into a recommendation model, wherein the recommendation model includes a behavior layer graph attention network, a graph attention network fused with explicit edge embedding, and a joint prediction module, and the behavior layer graph attention network includes a node-level multi-head attention network module and a behavior-level graph attention layer module fused with residual connection;

[0138] The second module 602 is used to perform node embedding processing on the multi-behavior heterogeneous graph through the node-level multi-head attention network module to obtain a behavior node embedding set;

[0139] The third module 603 is used to perform residual connection processing on the behavior node embedding set through the behavior-level graph attention layer module fused with residual connection to obtain a first node embedding set;

[0140] The fourth module 604 is used to perform edge embedding display adding processing on the multi-behavior heterogeneous graph through the graph attention network of the fused display edge embedding to obtain a second node embedding set;

[0141] A fifth module 605 is configured to perform joint prediction processing on the first node embedding set and the second node embedding set through the joint prediction module, and output a prediction score;

[0142] The sixth module 606 is used to perform recommendation processing according to the prediction score to obtain multiple behavior recommendation results.

[0143] It can be understood that the contents of the above method embodiments are all applicable to the present system embodiments, the functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0144] The embodiment of the present application also provides an electronic device, the electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the above-mentioned dual-channel perception multi-behavior recommendation method when executing the computer program. The electronic device can be any smart terminal including a tablet computer, a car computer, etc.

[0145] It can be understood that the contents of the above method embodiments are all applicable to the present device embodiments, the functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0146] See also Figure 7 , Figure 7 The hardware structure of an electronic device of another embodiment is illustrated, and the electronic device includes:

[0147] The processor 701 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application;

[0148] The memory 702 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 702 can store an operating system and other application programs. When the technical solution provided in the embodiment of this specification is implemented by software or firmware, the relevant program code is stored in the memory 702, and the processor 701 calls and executes the dual-channel perception multi-behavior recommendation method of the embodiment of this application;

[0149] Input / output interface 703, used to implement information input and output;

[0150] Communication interface 704, used to realize communication interaction between the device and other devices, which can be realized through wired mode (such as USB, network cable, etc.) or wireless mode (such as mobile network, WIFI, Bluetooth, etc.);

[0151] A bus 705 that transmits information between the various components of the device (e.g., the processor 701, the memory 702, the input / output interface 703, and the communication interface 704);

[0152] The processor 701 , the memory 702 , the input / output interface 703 and the communication interface 704 are connected to each other in communication within the device via a bus 705 .

[0153] An embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned dual-channel perception multi-behavior recommendation method.

[0154] It can be understood that the contents of the above method embodiments are all applicable to the present storage medium embodiments, the functions specifically implemented by the present storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0155] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0156] The embodiment of the present application provides a dual-channel perception multi-behavior recommendation method, system, device and medium. The scheme obtains a multi-behavior heterogeneous graph, inputs the multi-behavior heterogeneous graph into the recommendation model, performs node embedding processing on the multi-behavior heterogeneous graph through a node-level multi-head attention network module to obtain a behavior node embedding set; performs residual connection processing on the behavior node embedding set through a behavior-level graph attention layer module that integrates residual connections to obtain a first node embedding set; performs edge embedding display addition processing on the multi-behavior heterogeneous graph through a graph attention network that integrates display edge embedding to obtain a second node embedding set; performs joint prediction processing on the first node embedding set and the second node embedding set through a joint prediction module, and outputs a prediction score; performs recommendation processing based on the prediction score to obtain a multi-behavior recommendation result. The embodiment of the present application adopts a graph attention network to aggregate neighbor node information, and at the same time adopts a dual-channel strategy at the behavior level, using the behavior layer graph attention network and the graph attention network that integrates explicit edge embedding to obtain behavior features, which enables the recommendation model to obtain subtle differences in features between behaviors, significantly enhances the expressiveness of the model, and thus improves the efficiency and accuracy of recommendations.

[0157] The embodiments described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions provided in the embodiments of the present application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0158] Those skilled in the art will appreciate that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0159] The system embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment.

[0160] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0161] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0162] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0163] In the several embodiments provided in the present application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are only schematic. For example, the division of the above units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of systems or units, which can be electrical, mechanical or other forms.

[0164] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0165] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0166] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including multiple instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), disk or optical disk and other media that can store programs.

[0167] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present invention should be within the scope of the rights of the present invention.

Claims

1. A dual-channel perception multi-behavior recommendation method, characterized in that: The method comprises the following steps: Acquire a multi-behavior heterogeneous graph, and input the multi-behavior heterogeneous graph into a recommendation model, wherein the recommendation model includes a behavior-layer graph attention network, a graph attention network fused with explicit edge embedding, and a joint prediction module, and the behavior-layer graph attention network includes a node-level multi-head attention network module and a behavior-level graph attention layer module fused with residual connections; Performing node embedding processing on the multi-behavior heterogeneous graph through the node-level multi-head attention network module to obtain a behavior node embedding set; Performing residual connection processing on the behavior node embedding set through the behavior-level graph attention layer module fused with residual connection to obtain a first node embedding set; Performing edge embedding display adding processing on the multi-behavior heterogeneous graph through the graph attention network of the fused display edge embedding to obtain a second node embedding set; Performing joint prediction processing on the first node embedding set and the second node embedding set by the joint prediction module, and outputting a prediction score; Recommendation processing is performed according to the prediction scores to obtain multiple behavior recommendation results.

2. The method according to claim 1, characterized in that: The node-level multi-head attention network module is used to perform node embedding processing on the multi-behavior heterogeneous graph to obtain a behavior node embedding set, including the following steps: Performing high-dimensional feature mapping and attention calculation processing on the multi-behavior heterogeneous graph to obtain initial node embedding and initial neighbor node weights; Performing nonlinear transformation and normalization processing on the initial neighbor node weights to obtain target neighbor node weights; The initial node embedding is updated and multi-head attention calculation is performed according to the target neighbor node weight to obtain the behavior node embedding set.

3. The method according to claim 1, characterized in that The method of performing residual connection processing on the behavior node embedding set by the behavior-level graph attention layer module fused with residual connection to obtain a first node embedding set includes the following steps: Performing multi-layer perception and nonlinear transformation processing on the behavior node embedding set to obtain initial behavior node weights; Normalizing the initial behavior node weight to obtain a target behavior node weight; Aggregating the behavior node embedding set according to the target behavior node weight to obtain an aggregated embedding set; Perform residual calculation processing on the aggregated embedding set according to the residual parameter to obtain the first node embedding set.

4. The method according to claim 1, characterized in that: The step of performing edge embedding display adding processing on the multi-behavior heterogeneous graph through the graph attention network of the fused display edge embedding to obtain a second node embedding set includes the following steps: Performing behavior embedding and graph attention calculation processing on the multi-behavior heterogeneous graph to obtain an initial attention score; Performing nonlinear transformation and normalization processing on the initial attention score to obtain a target attention score; Update the node embeddings of the multi-behavior heterogeneous graph according to the target attention scores to obtain an edge embedding set; Perform multi-head attention calculation and residual connection processing on the edge embedding set to obtain the second node embedding set.

5. The method according to claim 1, characterized in that The step of performing joint prediction processing on the first node embedding set and the second node embedding set by the joint prediction module and outputting a prediction score comprises the following steps: Performing inner product calculation processing of user node embedding and item node embedding on the first node embedding set to obtain a first score; Performing inner product calculation processing of the user node embedding and the item node embedding on the second node embedding set to obtain a second score; The first score and the second score are combined according to the learnable parameter to obtain the predicted score.

6. The method according to any one of claims 1 to 5, characterized in that: Before inputting the multi-behavior heterogeneous graph into the recommendation model, the method further includes pre-training the recommendation model, specifically including the following steps: Obtain positive sample data and negative sample data; Inputting the positive sample data and the negative sample data into the recommendation model for interactive prediction respectively, to obtain a positive sample prediction score and a negative sample prediction score; Performing loss processing on the positive sample prediction score and the negative sample prediction score by using a loss function to obtain a loss value; The parameters of the recommendation model are updated according to the loss value to obtain the trained recommendation model.

7. The method according to any one of claims 1 to 5, characterized in that: The obtaining of a multi-behavior heterogeneous graph comprises the following steps: Get user-project interaction records; Behavior analysis and graph construction processing are performed on the user-project interaction records to obtain the multi-behavior heterogeneous graph.

8. A dual-channel perception multi-behavior recommendation system, characterized in that: The system comprises: The first module is used to obtain a multi-behavior heterogeneous graph and input the multi-behavior heterogeneous graph into a recommendation model, wherein the recommendation model includes a behavior layer graph attention network, a graph attention network fused with explicit edge embedding, and a joint prediction module, and the behavior layer graph attention network includes a node-level multi-head attention network module and a behavior-level graph attention layer module fused with residual connection; The second module is used to perform node embedding processing on the multi-behavior heterogeneous graph through the node-level multi-head attention network module to obtain a behavior node embedding set; A third module is used to perform residual connection processing on the behavior node embedding set through the behavior-level graph attention layer module fused with residual connection to obtain a first node embedding set; A fourth module is used to perform edge embedding display adding processing on the multi-behavior heterogeneous graph through the graph attention network of the fused display edge embedding to obtain a second node embedding set; A fifth module, configured to perform joint prediction processing on the first node embedding set and the second node embedding set through the joint prediction module, and output a prediction score; The sixth module is used to perform recommendation processing according to the prediction score to obtain multiple behavior recommendation results.

9. An electronic device, characterized in that: The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor implements the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images