A multi-view social recommendation method based on adaptive relationship graph

Abstract

This invention discloses a multi-view social recommendation method based on adaptive relationship graphs. It constructs different types of relationship graphs; uses a graph convolutional neural network to propagate node features layer by layer in the relationship graphs; after multi-layer graph convolutional propagation, it obtains user node embeddings and item node embeddings that integrate user historical interaction information, and constructs an adaptive relationship graph based on prototype guidance; performs information propagation and feature aggregation on the relationship graphs and the adaptive relationship graphs, concatenating and enhancing the node representations learned under different structural views to obtain enhanced user node embeddings and item node embeddings; constructs a recommendation model, inputs the enhanced user node embeddings and item node embeddings into the recommendation model, calculates the difference in predicted preferences of users for positive and negative sample items, and ranks candidate items according to the difference in predicted preferences of users for positive and negative sample items to generate the final recommendation result.

Description

Technical Field

[0001] This invention relates to the field of intelligent recommendation technology for information processing machines, and in particular to a multi-view social recommendation method based on adaptive relationship graphs. Background Technology

[0002] Recommendation systems are widely used in e-commerce platforms, social media platforms, and online content platforms to recommend potentially interesting items to users based on their historical behavior.

[0003] In social recommendation scenarios, user preferences are not only derived from user-item interaction behavior, but are also influenced by social relationships between users and semantic associations between items.

[0004] By incorporating social relationship information and item semantic information, the recommendation system's ability to characterize user interests can be improved to some extent.

[0005] Existing social recommendation methods are typically based on graph representation learning techniques, which model users and items by constructing user-item interaction graphs, user-user relationship graphs, and item-item relationship graphs.

[0006] Related methods utilize graph neural networks to perform information propagation and feature aggregation on graph structures, thereby learning low-dimensional representations of nodes and capturing high-order collaborative relationships between users and items.

[0007] In the prior art, most methods rely on pre-built fixed graph structures for representation learning.

[0008] During model training, the adjacency relationships between nodes usually remain unchanged, making it difficult to reflect changes in user interests and the potential changes in the associated structure brought about by the evolution of semantic relationships of items.

[0009] Some methods dynamically construct relationship graphs by calculating node similarity. These methods require large-scale similarity calculations between nodes, resulting in high computational complexity and insufficient computational efficiency in large-scale data scenarios.

[0010] Existing methods for multi-source information fusion typically involve simple combination or unified modeling of user-item interaction information, social relationship information, and item semantic information. However, different information sources exhibit significant differences in data distribution, connection density, and noise levels, making direct fusion insufficient to fully leverage the complementary roles of various information sources.

[0011] Recent research needs to develop recommendation methods that can adaptively mine the potential relationship structures between users and items while ensuring computational efficiency, and effectively fuse multi-source information. Summary of the Invention

[0012] The purpose of this paper is to address the problems of existing technologies by providing a multi-view social recommendation method based on adaptive relationship graphs.

[0013] The technical solution adopted to achieve the purpose of this invention is: A multi-view social recommendation method based on adaptive relationship graphs includes the following steps: Step 1: Collect user-item interaction data and user relationship data from the recommendation dataset, and preprocess them. Based on the preprocessed data, represent users and items as nodes in the graph, and represent the interaction or association relationships between nodes as edges in the graph, thus constructing different types of relationship graphs. Step 2: A graph convolutional neural network is used to propagate the node features in the relationship graph constructed in Step 1 layer by layer. After multiple layers of graph convolutional propagation, the user node embedding representation and the item node embedding representation are obtained by integrating user historical interaction information. Step 3: Based on the user node embedding representation and item node embedding representation obtained in Step 2, construct an adaptive relationship graph based on prototype guidance. Step 4: Information propagation and feature aggregation are performed on the relationship graph constructed in Step 1 and the adaptive relationship graph constructed in Step 3. The node representations learned under different structural views are spliced ​​and enhanced to obtain the enhanced user node embedding representation and item node embedding representation. Step 5: Construct a recommendation model. Input the enhanced user node embedding representation and item node embedding representation from Step 4 into the recommendation model, calculate the difference in predicted preferences of users for positive and negative sample items, rank the candidate items according to the difference in predicted preferences of users for positive and negative sample items, and generate the final recommendation result.

[0014] In the above technical solutions, the different types of relationship diagrams include three diagram structures: user-item interaction diagram, user-user relationship diagram, and item-item relationship diagram. The user-item interaction graph is formed by adding edges between user-item pairs that have interacted, based on the historical interaction records between users and items. The user-user relationship graph is constructed by adding edges between user nodes that have social relationships, based on users' social relationship data. The item-item relationship graph calculates the similarity between items based on their attributes, categories, tags, or co-occurrence relationships in user behavior. It then selects item pairs with high similarity to add edges and constructs a semantic graph.

[0015] In the above technical solution, step 2 includes the following steps: S201: Initialize the embedding vectors of user nodes and item nodes, and based on the relationship graph constructed in step 1, perform weighted aggregation of the features of nodes and their neighboring nodes through adjacency relationships in each layer, and obtain new node representations through nonlinear mapping; S202: By iteratively updating the node features through a multi-layer graph neural network, user nodes can aggregate the feature information of the items they interact with, and item nodes can aggregate the feature information of the users they interact with, thus learning user embedding representations and item embedding representations that contain user behavior preference information.

[0016] In the above technical solution, step 3 includes the following steps: S301: Local neighbor sampling is performed in the training batch based on the similarity relationship between node embeddings; The local neighbor sampling includes: Calculate the similarity between each node and other nodes separately. For each node, select based on the similarity value. Each node is selected as a local neighbor candidate and stored in the candidate node pool. Random sampling from the candidate node pool Selected from Each node constructs a local neighbor set. ; The formula for calculating the similarity between each node and other nodes is as follows:

[0017] In the formula, Representative node With nodes The similarity between them; , Representing nodes respectively With nodes The embedding vector; S302: Introduce prototype vectors for global neighbor sampling; S303: Jointly model the sampled local neighbors and global neighbors to obtain the final neighbor set, and construct an adaptive relationship graph based on the final neighbor set to obtain the adjacency matrix of the adaptive relationship graph; The adaptive relationship graph includes a user-user adaptive relationship graph. And item-item adaptive relationship graph .

[0018] In the above technical solution, step 4 includes the following steps: S401: Information propagation is performed on the relational graph constructed in step 1 and the adaptive relational graph constructed in step 3 respectively, to learn the node representation under different structural views; S402: A personalized feature fusion mechanism is used to integrate the node representations learned from different structural views, and the node representations from different structural views are feature-aggregated. The feature aggregation expression for the node representations under different structural views is as follows:

[0019] In the formula, This represents the nodes that the user has pieced together from different structural views; This represents the representation that the user learns from the user-item interaction graph; This represents the node representation that the user learns from the static relational graph. This represents the node representation that the user learns from the adaptive relationship graph. S403: Utilize a meta-network to learn node representations from different structural views after feature aggregation. As input to the meta-network, a corresponding mapping matrix is ​​generated based on the fusion features of each user. and The features of the relationship graph constructed in step 1 and the adaptive relationship graph constructed in step 3 are personalized and enhanced, and then fused with the node representation learned in the user-item interaction graph to obtain the final user embedding vector. The final user embedding vector is as follows:

[0020] In the formula, This represents the final user embedding representation after fusing multi-source graph structural information; This represents the representation that the user learns from the user-item interaction graph. The weighting parameter represents the proportion of contributions from the adjustment between static and adaptive relationships. This represents the node representation that the user learns from the static relational graph. , The mapping matrix represents the generation of the meta-network; This represents the node representation that the user learns from the adaptive relationship graph. Represents the projection matrix.

[0021] In the above technical solution, calculating the difference in user's predicted preference for positive and negative sample items includes: S501: Calculate the user's predicted preference score for positive and negative sample items based on the inner product of the user embedding vector and the item embedding vector; The formula for calculating the user's predicted preference score for positive and negative sample items is as follows:

[0022] In the formula, On behalf of users For items Predicted preference score On behalf of users Node embedding representation; Representative items Node embedding representation; S502: Calculate the difference in predicted preferences between users for positive and negative sample items based on the user's predicted preference scores for positive and negative sample items; S503: Use the Sigmoid function to probabilistically model the difference in the user's predicted preferences for positive and negative sample items, and maximize the ranking probability of positive samples relative to negative samples.

[0023] In the above technical solution, the recommendation model uses a weighted approach to calculate the Bayesian personalized ranking loss. Compared with the total learning loss Joint optimization is performed, as shown in the following expression:

[0024] In the formula, This represents the overall optimization objective function of the recommendation model; Represents the Bayesian personalized ranking loss. The representative factor is used to balance the ranking optimization objective with the representation consistency constraint. Represents overall comparative learning constraints; The Bayesian personalized ranking loss function is expressed as follows:

[0025] In the formula, Represents the Bayesian personalized ranking loss; , σ represents the inner product of the user and item embedding vectors; σ(·) is the Sigmoid function; Represents user nodes; j represents positive sample items; j represents negative sample items; The expression for the total loss of contrastive learning is as follows:

[0026] In the formula, Compare the losses of representative user nodes; Compare the losses to the representative item nodes; Represents overall comparative learning constraints; For user nodes The contrastive loss expression is as follows:

[0027] In the formula, Represents the similarity function. Represents temperature parameter, This represents the normalized node representation; Representing user nodes Normalized embedding representation in interactive views; Representing user nodes Normalized embedding representation in the structural view; In the representative structure view, excluding the user Other user nodes The normalized embedding representation.

[0028] Another aspect of the present invention provides an electronic device comprising: one or more processors; and a memory for storing one or more programs, wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the multi-view social recommendation method based on adaptive relationship graphs as described above.

[0029] Another aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the multi-view social recommendation method based on adaptive relationship graphs as described above.

[0030] Another aspect of the present invention provides a computer program product comprising computer-executable instructions which, when executed, are used to implement the multi-view social recommendation method based on adaptive relationship graphs as described above.

[0031] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention introduces prototype vectors in the node embedding space to represent the global semantic structure and establishes potential node connections by utilizing the association between nodes and prototypes. This enables nodes to form global semantic associations by sharing semantic prototypes. Based on this, an adaptive relationship graph structure is constructed by combining local neighbor relationships, thereby improving the ability to characterize the potential relationship structure of nodes.

[0032] 2. This invention learns node representations by fusing structural information from the original relation graph and the adaptive relation graph. While maintaining the ability to model global semantic relationships, it reduces the calculation of pairwise similarity between nodes, lowers the computational complexity of the model, and improves the quality of node representations, thereby improving the recommendation accuracy and stability of the recommendation system. Attached Figure Description

[0033] Figure 1 The diagram shown is a flowchart of the multi-view social recommendation method based on adaptive relationship graphs according to the present invention.

[0034] Figure 2 The diagram shown is a flowchart of the adaptive relationship graph construction process described in this invention. Detailed Implementation

[0035] The present invention will be further described in detail below with reference to specific embodiments.

[0036] Example 1

[0037] A multi-view social recommendation method based on adaptive relationship graphs, see [link to relevant documentation]. Figure 1 This includes the following steps: Step 1: Collect user-item interaction data and user relationship data from the recommendation dataset, and preprocess them. Based on the preprocessed data, represent users and items as nodes in the graph, and represent the interaction or association relationships between nodes as edges in the graph, thereby constructing different types of relationship graphs.

[0038] The constructed relationship graph is randomly divided into training, validation, and test sets in an 8:1:1 ratio to ensure that the distribution of users and items in the three sets is consistent, and random seeds are recorded to ensure the reproducibility of the experiment.

[0039] Preprocessing of the user-item interaction data and user relationship data in the collected recommendation dataset includes: removing missing values, detecting outliers, and normalizing the data.

[0040] Different types of relationship diagrams include three diagram structures: user-item interaction diagrams, user-user relationship diagrams, and item-item relationship diagrams. The user-item interaction graph is formed by adding edges between user-item pairs that have interacted, based on the historical interaction records between users and items, and at the same time, the interaction matrix is ​​recorded.

[0041] The user-user relationship graph is constructed by adding edges between user nodes that have social relationships, based on users' social relationship data (such as friends and following lists).

[0042] The item-item relationship graph calculates the similarity between items based on their attributes, categories, tags, or co-occurrence relationships in user behavior. It then selects item pairs with high similarity to add edges and constructs a semantic graph.

[0043] Step 2: A graph convolutional neural network is used to propagate the node features in the relationship graph constructed in Step 1 layer by layer. After multi-layer graph convolutional propagation, the user node embedding representation and the item node embedding representation are obtained by integrating the user's historical interaction information.

[0044] The layer-by-layer propagation expression is as follows:

[0045] In the formula, Representing the The layer's output node is embedded in the representation; Represents a non-linear activation function; Represents a symmetric normalized adjacency matrix; Representing the Layer node representation; Represents the learnable weight matrix; This represents the number of layers in the graph convolutional network.

[0046] Step 2 includes the following steps: S201: Initialize the embedding vectors of user nodes and item nodes, and based on the relationship graph constructed in step 1, perform weighted aggregation of the features of nodes and their neighboring nodes through adjacency relationships in each layer, and obtain new node representations through nonlinear mapping.

[0047] S202: By iteratively updating the node features through a multi-layer graph neural network, user nodes can aggregate the feature information of the items they interact with, and item nodes can aggregate the feature information of the users they interact with, thereby learning user embedding representations and item embedding representations that contain user behavior preference information.

[0048] Step 3: Based on the user node embedding representation and item node embedding representation obtained in Step 2, construct an adaptive relationship graph based on prototype guidance.

[0049] The adaptive relationship graph includes a user-user adaptive relationship graph. And item-item adaptive relationship graph .

[0050] See Figure 2 Step 3 includes the following steps: S301: Local neighbor sampling is performed in the training batch based on the similarity relationship between node embedding representations.

[0051] That is, calculate the similarity between each node and other nodes separately, and for each node, select based on the similarity value. Each node is selected as a local neighbor candidate and stored in the candidate node pool. Random sampling from the candidate node pool Selected from Each node constructs a local neighbor set. To enhance the diversity and stability of the neighborhood.

[0052] The formula for calculating the similarity between each node and other nodes is as follows:

[0053] In the formula, Representative node With nodes The similarity between them; , Representing nodes respectively With nodes The embedding vector.

[0054] S302: Introduce prototype vectors for global neighbor sampling.

[0055] To characterize the global semantic structure in the embedding space, K prototype vectors are introduced to represent the global semantic structure in the embedding space. ( Modeling is performed on the representation of latent semantic patterns, enabling different nodes to establish latent semantic relationships through shared prototypes, thereby effectively supplementing the potential collaborative relationships between nodes; specifically including: calculating separately The similarity between each node and each prototype vector is used to select the most relevant node for each node. Each node has a set of prototypes that constitute its prototype association set. ,when When a node has a high similarity to the same prototype vector (the high similarity is a pre-set threshold), then it is considered... Each node has potential relationships in the semantic space. Based on shared prototype relationships, global neighbor relationships between nodes are sampled to construct a global neighbor set. .

[0056] For example, this embodiment presupposes n nodes and K prototype vectors. For each node, its relationship with the prototype vector needs to be calculated separately. The similarity between the prototype vectors, and the total number of operations for similarity calculation is:

[0057] In the formula, This represents the total number of operations performed in the similarity calculation. Represents the number of nodes; Represents the prototype vector.

[0058] The time complexity for each node is:

[0059] In the formula, Represents the number of prototype vectors, and usually satisfies ; This represents the asymptotic upper bound of the algorithm's time complexity.

[0060] The complexity of similarity calculation based on this embodiment is... Reduced to .

[0061] For example, when node and nodes satisfy If they are considered to have a potential relationship in the global semantic space, then the nodes are considered to be related. Add node The global candidate set.

[0062] The global neighbor set The candidate with the highest similarity selected from the global candidate set Each node.

[0063] Each prototype vector represents a potential global semantic pattern in the embedding space. By guiding the construction of relationships between nodes through prototype vectors, it is possible to avoid calculating pairwise similarity for all node pairs, thereby reducing the computational complexity in the relationship graph construction process and improving the scalability of the model in large-scale recommendation scenarios.

[0064] S303: Jointly model the sampled local neighbors and global neighbors to obtain the final neighbor set, and construct an adaptive relationship graph based on the final neighbor set to obtain the adjacency matrix of the adaptive relationship graph.

[0065] The final neighbor set expression is as follows:

[0066] In the formula, Representative node The final set of neighbors; Represents a local set of neighbors; It represents the global set of neighbors.

[0067] For example, for nodes With nodes If node Belongs to node If the set of neighbors is found, then the corresponding position in the adjacency matrix is ​​assigned a value of 1; otherwise, it is assigned a value of 0. The expression for the adjacency matrix is ​​as follows:

[0068] In the formula, Representative node With nodes Is there a connection between them? Represents the index of any node in the graph; Representative node The final set of neighbors.

[0069] Step 4: Information propagation and feature aggregation are performed on the relationship graph constructed in Step 1 and the adaptive relationship graph constructed in Step 3. The node representations learned under different structural views are spliced ​​and enhanced to obtain the enhanced user node embedding representation and item node embedding representation.

[0070] Among them, the The node update expression for the layer is as follows:

[0071] In the formula, The target node is at the Hierarchical relationship diagram representation; This represents the aggregation of neighbor features; Represents the relationship between the target node and its neighboring nodes. Perform feature encoding; The target node is at the Layer representation; Representing neighbor nodes In the Layer representation; The target node to be updated in the relational graph; Represents the target node The neighboring nodes.

[0072] Step 4 includes the following steps: S401: Information propagation is performed on the relational graph constructed in step 1 and the adaptive relational graph constructed in step 3 respectively, to learn the node representation under different structural views.

[0073] S402: The personalized feature fusion mechanism is used to integrate the node representations learned under different structural views, and the node representations under different structural views are aggregated for features (the nodes learned under different structural views are uniformly represented, so that the nodes after uniform representation form multi-source feature input (multi-view representation) without losing the original semantic information).

[0074] The user node embedding representation and item node embedding representation after feature aggregation include the user node embedding representation under adaptive relationships. and item node embedding representation And user node embedding representation under static relationships and item node embedding representation .

[0075] The feature aggregation expression for the node representations under different structural views is as follows:

[0076] In the formula, This represents the nodes that the user has pieced together from different structural views; This represents the representation that the user learns from the user-item interaction graph. This represents the node representation that the user learns from the static relational graph. This represents the node representation that the user learns from the adaptive relational graph.

[0077] S403: Utilize a meta-network (fully connected network MLP) to learn node representations from different structural views after feature aggregation. As input to the meta-network, a corresponding mapping matrix is ​​generated based on the fusion features of each user. and The features of the relationship graph constructed in step 1 and the adaptive relationship graph constructed in step 3 are personalized and enhanced, and then fused with the node representation learned in the user-item interaction graph to obtain the final user embedding vector.

[0078] The final user embedding vector is as follows:

[0079] In the formula, This represents the final user embedding representation after fusing multi-source graph structural information; This represents the representation that the user learns from the user-item interaction graph. The weighting parameter represents the proportion of contributions from the adjustment between static and adaptive relationships. This represents the node representation that the user learns from the static relational graph. , The mapping matrix represents the generation of the meta-network; This represents the node representation that the user learns from the adaptive relationship graph. Represents the projection matrix.

[0080] Step 5: Construct a recommendation model. Input the enhanced user node embedding representation and item node embedding representation from Step 4 into the recommendation model to obtain the user's predicted preference scores for positive and negative sample items. Calculate the user's predicted preference difference for positive and negative sample items based on the user's predicted preference scores for positive and negative sample items. Sort the candidate items according to the user's predicted preference difference for positive and negative sample items to generate the final recommendation result (personalized recommendation list).

[0081] The formula for calculating the user's predicted preference score for positive and negative sample items is as follows:

[0082] In the formula, On behalf of users For items Predicted preference score On behalf of users Node embedding representation; Representative items The node embedding representation.

[0083] The recommendation model in this embodiment uses Bayesian Personalized Ranking (BPR) loss as the main optimization objective function, and constructs a triplet of user-positive sample-negative sample (where positive sample items...). Indicates user Items that have actually been interacted with in historical interactions are negative sample items. (Representing items for which the user did not interact), calculate the difference in the user's predicted preferences for positive and negative sample items. Furthermore, the Sigmoid function is used to model the difference in users' predicted preferences for positive and negative sample items, thereby minimizing the Bayesian personalized ranking loss function.

[0084] The difference in predicted preferences between users for positive and negative sample items essentially stems from the predicted preference scores of users for positive and negative samples. It is a relative comparison used to characterize the ranking relationship between users for different items, rather than the absolute preference for a single item. The BPR loss is optimized based on this difference, so that the predicted score of positive samples is higher than that of negative samples, thereby improving the ranking performance of the recommendation results.

[0085] The difference in users' prediction preferences for positive and negative sample items The calculations include: S501: Calculate the user's predicted preference score for positive and negative sample items based on the inner product of the user's embedding vector and the item's embedding vector. .

[0086] S502: Calculate the difference in predicted preferences between users for positive and negative sample items based on the user's predicted preference scores for positive and negative sample items. The difference in user prediction preferences for positive and negative sample items reflects the degree to which the model indicates that users are more inclined to choose positive sample items than negative sample items.

[0087] S503: Utilize the Sigmoid function to analyze the user's predicted preference difference for positive and negative sample items. Probabilistic modeling is performed, and the BPR loss function is minimized by maximizing the ranking probability of positive samples relative to negative samples, thereby optimizing and learning the recommendation model parameters.

[0088] The Bayesian personalized ranking loss function expression is as follows:

[0089] In the formula, Represents the Bayesian personalized ranking loss; , σ represents the inner product of the user and item embedding vectors; σ(·) is the Sigmoid function; Represents user nodes; j represents positive sample items; j represents negative sample items.

[0090] To enhance the consistency of representation across different views, a contrastive learning mechanism is introduced to align the representation of the same node in the interactive view (UI) and the structure view (S).

[0091] The interactive view consists of user-item interaction relationships, and the structural view consists of user-user relationships and item-item relationships.

[0092] For any user node Each user node is obtained. In interactive views and structure views, this is denoted as... and .

[0093] According to user node The representations in the interactive view and the structure view are used, and a node-level contrastive loss is constructed using an InfoNCE-based contrastive learning mechanism.

[0094] For user nodes The contrastive loss expression is as follows:

[0095] In the formula, Represents the similarity function. Represents temperature parameter, This represents the normalized node representation; Representing user nodes Normalized embedding representation in interactive views; Representing user nodes Normalized embedding representation in the structural view; In the representative structure view, excluding the user Other user nodes The normalized embedding representation.

[0096] Similarly, item node comparison loss Constructed in the same way.

[0097] Therefore, the expression for the total loss of contrastive learning is as follows:

[0098] In the formula, Compare the losses of representative user nodes; Compare the losses to the representative item nodes; This represents the overall comparative learning constraint.

[0099] The recommendation model in this embodiment uses a weighted approach to evaluate the Bayesian personalized ranking loss. Compared with the total learning loss Joint optimization is performed, as shown in the following expression:

[0100] In the formula, This represents the overall optimization objective function of the recommendation model; Represents the Bayesian personalized ranking loss. The representative factor is used to balance the ranking optimization objective with the representation consistency constraint. This represents the overall comparative learning constraint.

[0101] During the training of the recommendation model, each loss term is jointly optimized in a unified parameter space. Through backpropagation, the parameters of the graph convolutional network, node embedding representation, and feature fusion module are jointly updated, thereby achieving a synergistic improvement in recommendation performance and representation learning ability.

[0102] The embedding vector dimension of the recommendation model is fixed at 64. The Adam optimizer is used for model training. The temperature coefficient for contrastive learning is set to 0.6, the initial learning rate is 1e-2, and the number of graph convolutional layers is set to 2.

[0103] Example 2

[0104] The present invention also provides a computer-readable storage medium carrying one or more programs, which, when executed, implement the multi-view social recommendation method based on adaptive relationship graphs according to Embodiment 1 of the present invention.

[0105] According to embodiments of the present invention, the computer-readable storage medium may be a non-volatile computer-readable storage medium.

[0106] Examples may include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0107] In this invention, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0108] Example 3

[0109] Embodiments of the present invention also include a computer program product comprising a computer program containing program code for performing the methods provided in the embodiments of the present invention. When the computer program product is run on an electronic device, the program code is used to enable the electronic device to implement the multi-view social recommendation method based on adaptive relationship graphs as described in Embodiment 1 of the present invention.

[0110] For ease of explanation, spatial relative terms such as "upper," "lower," "left," and "right" are used in the embodiments to describe the relationship of one element or feature shown in the figure relative to another element or feature.

[0111] It should be understood that, in addition to the orientations shown in the diagram, spatial terms are intended to include different orientations of the device during use or operation.

[0112] For example, if the device in the figure is inverted, the element described as being "below" other elements or features will be positioned "above" other elements or features.

[0113] Therefore, the exemplary term "below" can include both "above" and "below".

[0114] The device can be positioned in other ways (rotated 90 degrees or in other orientations), and the spatial relative description used here can be interpreted accordingly.

[0115] Moreover, relational terms such as “first” and “second” are used merely to distinguish one component from another that has the same name, without necessarily requiring or implying any such actual relationship or order between the components.

[0116] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-view social recommendation method based on adaptive relationship graphs, characterized in that, Includes the following steps: Step 1: Collect user-item interaction data and user relationship data from the recommendation dataset, and preprocess them. Based on the preprocessed data, represent users and items as nodes in the graph, and represent the interaction or association relationships between nodes as edges in the graph, thus constructing different types of relationship graphs. Step 2: A graph convolutional neural network is used to propagate the node features in the relationship graph constructed in Step 1 layer by layer. After multiple layers of graph convolutional propagation, the user node embedding representation and the item node embedding representation are obtained by integrating user historical interaction information. Step 3: Based on the user node embedding representation and item node embedding representation obtained in Step 2, construct an adaptive relationship graph based on prototype guidance. Step 4: Information propagation and feature aggregation are performed on the relationship graph constructed in Step 1 and the adaptive relationship graph constructed in Step 3. The node representations learned under different structural views are spliced ​​and enhanced to obtain the enhanced user node embedding representation and item node embedding representation. Step 5: Construct a recommendation model. Input the enhanced user node embedding representation and item node embedding representation from Step 4 into the recommendation model, calculate the difference in predicted preferences of users for positive and negative sample items, rank the candidate items according to the difference in predicted preferences of users for positive and negative sample items, and generate the final recommendation result.

2. The multi-view social recommendation method according to claim 1, characterized in that, Different types of relationship diagrams include three diagram structures: user-item interaction diagrams, user-user relationship diagrams, and item-item relationship diagrams. The user-item interaction graph is formed by adding edges between user-item pairs that have interacted, based on the historical interaction records between users and items. The user-user relationship graph is constructed by adding edges between user nodes that have social relationships, based on users' social relationship data. The item-item relationship graph calculates the similarity between items based on their attributes, categories, tags, or co-occurrence relationships in user behavior. It then selects item pairs with high similarity to add edges and constructs a semantic graph.

3. The multi-view social recommendation method according to claim 1, characterized in that, Step 2 includes the following steps: S201: Initialize the embedding vectors of user nodes and item nodes, and based on the relationship graph constructed in step 1, perform weighted aggregation of the features of nodes and their neighboring nodes through adjacency relationships in each layer, and obtain new node representations through nonlinear mapping; S202: By iteratively updating the node features through a multi-layer graph neural network, user nodes can aggregate the feature information of the items they interact with, and item nodes can aggregate the feature information of the users they interact with, thus learning user embedding representations and item embedding representations that contain user behavior preference information.

4. The multi-view social recommendation method according to claim 1, characterized in that, Step 3 includes the following steps: S301: Local neighbor sampling is performed in the training batch based on the similarity relationship between node embeddings; The local neighbor sampling includes: Calculate the similarity between each node and other nodes separately. For each node, select based on the similarity value. Each node is selected as a local neighbor candidate and stored in the candidate node pool. Random sampling from the candidate node pool Selected from Each node constructs a local neighbor set. ; The formula for calculating the similarity between each node and other nodes is as follows: In the formula, Representative node With nodes The similarity between them; , Representing nodes respectively With nodes The embedding vector; S302: Introduce prototype vectors for global neighbor sampling; S303: Jointly model the sampled local neighbors and global neighbors to obtain the final neighbor set, and construct an adaptive relationship graph based on the final neighbor set to obtain the adjacency matrix of the adaptive relationship graph; The adaptive relationship graph includes a user-user adaptive relationship graph. And item-item adaptive relationship graph .

5. The multi-view social recommendation method according to claim 1, characterized in that, Step 4 includes the following steps: S401: Information propagation is performed on the relational graph constructed in step 1 and the adaptive relational graph constructed in step 3 respectively, to learn the node representation under different structural views; S402: A personalized feature fusion mechanism is used to integrate the node representations learned from different structural views, and the node representations from different structural views are feature-aggregated. The feature aggregation expression for the node representations under different structural views is as follows: In the formula, This represents the nodes that the user has pieced together from different structural views; This represents the representation that the user learns from the user-item interaction graph. This represents the node representation that the user learns from the static relational graph. This represents the node representation that the user learns from the adaptive relationship graph. S403: Utilize a meta-network to learn node representations from different structural views after feature aggregation. As input to the meta-network, a corresponding mapping matrix is ​​generated based on the fusion features of each user. and The features of the relationship graph constructed in step 1 and the adaptive relationship graph constructed in step 3 are personalized and enhanced, and then fused with the node representation learned in the user-item interaction graph to obtain the final user embedding vector. The final user embedding vector is as follows: In the formula, This represents the final user embedding representation after fusing multi-source graph structural information; This represents the representation that the user learns from the user-item interaction graph. The weighting parameter represents the proportion of contributions from the adjustment between static and adaptive relationships. This represents the node representation that the user learns from the static relational graph. , The mapping matrix represents the generation of the meta-network; This represents the node representation that the user learns from the adaptive relationship graph. Represents the projection matrix.

6. The multi-view social recommendation method according to claim 1, characterized in that, The calculation of the difference in user prediction preferences for positive and negative sample items includes: S501: Calculate the user's predicted preference score for positive and negative sample items based on the inner product of the user embedding vector and the item embedding vector; The formula for calculating the user's predicted preference score for positive and negative sample items is as follows: In the formula, On behalf of users For items Predicted preference score On behalf of users Node embedding representation; Representative items Node embedding representation; S502: Calculate the difference in predicted preferences between users for positive and negative sample items based on the user's predicted preference scores for positive and negative sample items; S503: Use the Sigmoid function to probabilistically model the difference in the user's predicted preferences for positive and negative sample items, and maximize the ranking probability of positive samples relative to negative samples.

7. The multi-view social recommendation method according to claim 1, characterized in that, The recommendation model employs a weighted approach to evaluate the Bayesian personalized ranking loss. Compared with the total learning loss Joint optimization is performed, as shown in the following expression: In the formula, This represents the overall optimization objective function of the recommendation model; Represents the Bayesian personalized ranking loss. The representative factor is used to balance the ranking optimization objective with the representation consistency constraint. Represents overall comparative learning constraints; The Bayesian personalized ranking loss function is expressed as follows: In the formula, Represents the Bayesian personalized ranking loss; , σ represents the inner product of the user and item embedding vectors; σ(·) is the Sigmoid function; Represents user nodes; j represents positive sample items; j represents negative sample items; The expression for the total loss of contrastive learning is as follows: In the formula, Compare the losses of representative user nodes; Compare the losses to the representative item nodes; Represents overall comparative learning constraints; For user nodes The contrastive loss expression is as follows: In the formula, Represents the similarity function. Represents temperature parameter, This represents the normalized node representation; Representing user nodes Normalized embedding representation in interactive views; Representing user nodes Normalized embedding representation in the structural view; In the representative structure view, excluding the user Other user nodes The normalized embedding representation.

8. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the multi-view social recommendation method based on adaptive relationship graphs as described in claim 1.

9. A computer-readable storage medium, characterized in that, It stores computer-executable instructions, which, when executed, are used to implement the multi-view social recommendation method based on adaptive relationship graphs as described in claim 1.

10. A computer program product, characterized in that, The aforementioned computer program product includes computer-executable instructions that, when executed, implement the multi-view social recommendation method based on adaptive relationship graphs as described in claim 1.

Images