A Personalized Recommendation Method Based on Multi-View Heterogeneous Attribute Graph Structure Learning

By employing a multi-view heterogeneous attribute graph structure learning method, subgraphs of heterogeneous recommendation graphs are decomposed and enhanced. Combined with higher-order connectivity and graph neural networks, the noise and fit problems in heterogeneous graph recommendation are solved, thereby improving recommendation accuracy and robustness.

CN122864601BActive Publication Date: 2026-11-17SHANDONG UNIV OF SCI & TECH +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611365663.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-04
Publication Date
2026-11-17
Estimated Expiration
2046-09-04

AI Technical Summary

Technical Problem

Existing recommendation methods based on heterogeneous graphs suffer from noise and missing information in the interaction graphs. Furthermore, existing graph structure learning methods lack adaptability and cannot effectively handle the multiple types of nodes and edges in heterogeneous graphs, leading to a decrease in recommendation accuracy.

Method used

A multi-view heterogeneous attribute graph structure learning method is adopted. By constructing a heterogeneous recommendation attribute graph, it is split into multiple relation subgraphs and attribute, semantic and topological enhancements are performed on each subgraph. The subgraphs are then fused using a multi-channel attention mechanism to generate a high-order connected graph, which is then combined with a graph neural network for end-to-end optimization.

Benefits of technology

It improves the integrity and reliability of heterogeneous recommendation graph structures, reduces noise connection interference, captures potential associations, and enhances recommendation accuracy, especially in the case of sparse interactive data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122864601B_ABST
    Figure CN122864601B_ABST
Patent Text Reader

Abstract

This invention provides a personalized recommendation method based on multi-view heterogeneous attribute graph structure learning, relating to the fields of graph data processing and intelligent recommendation technology. The specific steps are as follows: constructing a heterogeneous recommendation attribute graph; splitting the heterogeneous recommendation attribute graph into relation subgraphs, enhancing the relation subgraphs to generate three types of enhanced relation subgraphs; fusing the three types of enhanced relation subgraphs of each relation subgraph to obtain local enhanced graphs of each relation subgraph, and aggregating all local enhanced graphs to obtain a global enhanced graph; calculating a higher-order connected graph; fusing the global enhanced graph and the higher-order connected graph to generate an optimized recommendation graph; inputting the optimized recommendation graph into a graph neural network encoder to generate a low-dimensional embedding representation; constructing a joint loss function; and generating a Top-K recommendation list based on preference scores. The technical solution of this invention overcomes the problems of existing technologies that simplify the graph structure from only a single dimension, ignore higher-order connectivity, and adopt a separate two-stage approach to independently optimize the graph structure and recommendation embedding.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of graph data processing and intelligent recommendation technology, specifically to a personalized recommendation method based on learning the structure of multi-view heterogeneous attribute graphs. Background Technology

[0002] In recent years, graph neural networks (GNNs) have become one of the mainstream technologies for recommender systems due to their powerful ability to aggregate topological information: by constructing user-item interaction graphs, GNNs aggregate neighborhood information to generate embedded representations of users and items, thereby achieving preference prediction and personalized ranking.

[0003] In real-world recommendation scenarios, the interaction network is essentially a heterogeneous graph. However, existing recommendation methods based on heterogeneous graphs suffer from two major drawbacks: First, the interaction graph inherently contains noise and missing information: user misclicks and accidental favorites introduce noisy edges, while the interaction links corresponding to potential user interests remain missing. Second, GNNs rely on topological structures for information transmission, and noise and missing edges directly reduce embedding quality, leading to decreased recommendation accuracy. Third, existing Graph Structure Learning (GSL) methods lack adaptability: most GSL algorithms are designed for homogeneous graphs and cannot handle the diverse types of nodes and edges in heterogeneous graphs. A few heterogeneous graph structure learning methods only optimize the structure from attribute and semantic dimensions, lacking topological-dimensional structure enhancement strategies, resulting in limited robustness and generalization ability, and failing to directly adapt to the ranking optimization goals of recommendation scenarios.

[0004] Therefore, there is a need for a personalized recommendation method based on multi-view heterogeneous attribute graph structure learning that can simultaneously enhance graph structures from multiple dimensions such as attributes, semantics, and topology, integrate higher-order connectivity, and jointly optimize graph structures and recommendation embeddings end-to-end. Summary of the Invention

[0005] The main objective of this invention is to provide a personalized recommendation method based on multi-view heterogeneous attribute graph structure learning, in order to solve the problem in the prior art that simplifies the graph structure from only a single dimension, ignores higher-order connectivity, and adopts a separate two-stage approach to independently optimize the graph structure and recommendation embedding.

[0006] To achieve the above objectives, this invention provides a personalized recommendation method based on multi-view heterogeneous attribute graph structure learning, specifically including the following steps: S1. Obtain node data, node attribute data, and interaction relationship data between nodes in the recommendation scenario, construct a heterogeneous recommendation attribute graph, and establish a unified node index space according to all nodes in the heterogeneous recommendation attribute graph.

[0007] S2: The heterogeneous recommendation attribute graph is split into multiple relation subgraphs based on the edge type. Attribute enhancement, semantic enhancement, and topological enhancement are performed on each relation subgraph to generate three types of enhanced relation subgraphs.

[0008] S3 adaptively fuses the three types of enhanced relational subgraphs of each relational subgraph through a multi-channel attention mechanism to obtain the local enhanced graph of each relational subgraph. The local enhanced graphs of all relational subgraphs are then mapped to a unified node space and aggregated to obtain the global enhanced graph.

[0009] S4. Based on the adjacency matrix of the heterogeneous recommendation attribute graph in the unified node index space, calculate the node connectivity from first order to the preset highest order, and fuse the node connectivity of different orders to obtain a high-order connected graph.

[0010] S5 uses a multi-channel attention layer to adaptively fuse the global enhancement graph and the higher-order connected graph to generate the final optimized recommendation graph.

[0011] S6 optimizes the input graph neural network encoder of the recommendation graph, aggregating neighborhood information to generate low-dimensional embedded representations of user nodes and item nodes.

[0012] S7. Construct a joint loss function that includes recommendation ranking loss, graph structure reconstruction loss and structure regularization term, and backpropagate to iteratively optimize the recommendation graph parameters and graph neural network encoder parameters.

[0013] S8 calculates the target user's preference score for all candidate items based on the trained user and item embeddings, and generates a Top-K recommendation list after sorting.

[0014] Furthermore, the heterogeneous recommendation attribute graph in step S1 The definition is as follows: ; in, This represents a set of nodes, including: user nodes, merchant nodes, service category nodes, and rating level nodes; The set of edges includes: rating edges between users and merchants, friend edges between users, subordinate edges between merchants and service categories, and belonging edges between merchants and rating levels. This represents the set of types in the graph, including the set of node types. and edge type set ; A mapping function for node types; For edge type mapping functions; This is a matrix of node attributes.

[0015] Furthermore, in step S2, attribute enhancement is performed on each relation subgraph separately, specifically as follows: Based on the edge connection type, the heterogeneous recommendation attribute graph is split into multiple relation subgraphs, resulting in a set of relation subgraphs. : ; in, Indicates the first A relational subgraph, This represents the total number of relational subgraphs.

[0016] Construct a sparse attribute similarity graph, calculated using the following formula: ; in, The preset number of nearest neighbors; For the first The K-nearest neighbor attribute similarity adjacency matrix corresponding to each relation subgraph; Relational subgraph The attribute matrix of the nodes included in it; The graph construction operation is for K-nearest neighbors.

[0017] Multi-head weighted cosine similarity is used to calculate the attribute similarity between nodes. The calculation formula is: ; in, For the number of attention heads; For the first The dimension weight vector corresponding to the head; For Hadamah accumulation; The cosine similarity function; For nodes The original attribute feature vector; For nodes The original attribute feature vector.

[0018] A threshold is set on the similarity results for sparsity filtering, resulting in the feature similarity matrix of the head nodes of the relation subgraph. : ; in, The preset similarity threshold; and All are relational subgraphs The node index in the data.

[0019] Combining the subgraph adjacency matrix with forward and backward feature propagation yields two feature propagation graphs: ; in, For the first The head-end feature propagation graph corresponding to each relational subgraph; For the first Tail feature propagation graphs corresponding to each relational subgraph; Relational subgraph The adjacency matrix; and These are the feature similarity matrices of the head and tail nodes of the relation subgraph, respectively.

[0020] Perform a non-linear mapping on the node attributes to obtain the mapped node attributes. : ; in, For nodes The type of node it belongs to; , These are the trainable weight matrix and bias vector for the corresponding node type, respectively; It is the Sigmoid non-linear activation function.

[0021] Node similarity is calculated in the mapped latent space, and the latent space attribute similarity map is obtained after thresholding. : ; in, For the first The multiple weighted cosine similarity function of each relation subgraph in the latent attribute space; and They are nodes and nodes The latent attribute representation vector obtained after passing through the node type-specific mapping layer; This is the threshold for the similarity of potential spatial attributes.

[0022] The attribute similarity map, forward feature propagation map, backward feature propagation map, and latent space attribute similarity map are treated as four channels and adaptively weighted and fused through a multi-channel attention layer to obtain the final attribute enhancement map. : ; in, For multi-channel attention fusion function; For the first The attribute augmentation graph corresponding to each relation subgraph.

[0023] Furthermore, in step S2, semantic enhancement is performed on each relation subgraph, specifically as follows: Preset candidate metapath set : ; in, For the first Candidate metapaths.

[0024] The semantic similarity between nodes is calculated based on the candidate metapath node representation, and the semantic candidate graph is obtained after threshold filtering. : ; in, For nodes In the Semantic representation vectors under each candidate metapath; For nodes In the Semantic representation vectors under each candidate metapath; This is a similarity calculation function; This is the semantic similarity threshold.

[0025] The graph is adaptively merged into a final semantically enhanced graph using an attention mechanism. ; in, For semantic channel attention fusion function; For the first The semantic augmentation graph corresponding to each relation subgraph.

[0026] Furthermore, in step S2, topology enhancement is performed on each relational subgraph, specifically as follows: For relational subgraphs adjacency matrix First, self-loop addition and normalization are performed to obtain a normalized adjacency matrix; then, the topological correlation between nodes is calculated based on the diffusion kernel function, random walk transition matrix, or personalized PageRank. ; ; in, This is the degree diagonal matrix of the correspondence subgraph; It is the identity matrix; Represents the adjacency matrix of the relational subgraph; For the current node; The adjacent nodes corresponding to the current node.

[0027] The topology correlation matrix is ​​thresholded, Top-K truncated, normalized, or sparsified to form a topology enhancement graph: ; in, Indicates the first The topology enhancement graph corresponding to each relational subgraph; This is the restart probability hyperparameter.

[0028] Furthermore, step S3 specifically includes the following steps: S3.1, for each relational subgraph The attribute-enhanced graph, semantic-enhanced graph, and topological-enhanced graph are input into a multi-channel attention layer as different structural channels. The multi-channel attention layer assigns weights based on the contribution of the attribute-enhanced graph, semantic-enhanced graph, and topological-enhanced graph to the current relation subgraph, thus obtaining the local enhancement graph corresponding to the relation subgraph. ; in, Indicates the first The first attention fusion function for each relational subgraph; Indicates the first Local enhancement maps corresponding to each relational subgraph.

[0029] S3.2, Local Enhancement Graph of All Relational Subgraphs Mapping to a unified node space and aggregating them yields a global augmented graph. : ; in, This is a global enhancement graph.

[0030] Furthermore, the specific calculation method for the higher-order connected graph in step S4 is as follows: Constructing high-order connected graphs based on first- to c-order connectivity calculations using adjacency matrices: ; in, Represents the adjacency matrix; Indicates a second-order connectivity relation; express Order-order connectivity; It is a high-order connected graph.

[0031] Normalize, truncate, sparsify, or adjust the order weights of each matrix, and set weights for different orders to construct the following weighted high-order connected graph. : ; in, , to This represents the weights corresponding to different orders.

[0032] Furthermore, the specific calculation method for optimizing the recommendation graph in step S5 is as follows: The global augmentation graph and the higher-order connected graph are adaptively weighted and merged to output the optimized recommendation graph used in the graph neural network encoder. : ; in, This represents the second attention fusion function; The corresponding adjacency matrix is ​​denoted as .

[0033] Furthermore, step S6 specifically includes the following steps: S6.1, normalize the adjacency matrix of the optimized recommendation graph: ; in, To optimize the adjacency matrix of the recommendation graph The corresponding degree diagonal matrix; This is the normalized adjacency matrix.

[0034] S6.2 uses a graph convolutional network encoder for encoding. The encoding formula of the two-layer graph convolutional network encoder (GCN) is as follows: ; in, This is the first layer of trainable weight matrix; For input attribute dimensions; For the hidden layer dimension; This is the trainable weight matrix for the second layer; Embed dimensions for preset nodes; It is a non-linear activation function; This is the final output matrix representing the nodes.

[0035] Furthermore, the specific calculation method for the joint loss function in step S7 is as follows: The recommendation ranking supervision loss adopts Bayesian personalized ranking BPR loss. The specific definition is as follows: ; in, For target user nodes; These are positive sample products; These are negative sample products; To train the set of triples; To optimize the recommended image User obtained through encoding The embedding vector; and Positive sample products and negative sample products The embedding vector; This is the set of all trainable weight matrices for a graph neural network encoder. This is the weight decay coefficient; is the square of the Frobenius norm of the matrix; This is a transpose operation; It is a logarithmic function.

[0036] Graph structure reconstruction loss The specific form is as follows: ; in, For nodes Optimize the recommendation image The set of direct neighbor nodes in the; For nodes The set of negative sampling nodes; For indicator functions; For nodes Optimize the recommendation image The embedding vector obtained from learning; Candidate nodes Optimize the recommendation image The embedding vector obtained from learning.

[0037] Regularization loss The specific formula is as follows: ; in, For a set of nodes The total number of nodes included in it; The adjacency matrix is ​​the first... Line number The element values ​​of the column; Let L1 be the norm of the matrix; This is the scalar absolute value operator.

[0038] Joint loss function for: ; in, For graph structure reconstruction loss Weighting coefficients; For structural regularization loss The weighting coefficients.

[0039] The present invention has the following beneficial effects: (1) According to different edge types, the present invention divides the heterogeneous recommendation attribute graph into multiple relation subgraphs and enhances the structure of each relation subgraph from the three views of attribute, semantic and topology. This can preserve the specific semantics included in different connection relationships, supplement the missing node associations in the heterogeneous recommendation attribute graph, reduce the interference of noisy connections on the propagation of neighborhood information, and thus improve the integrity and reliability of the heterogeneous recommendation graph structure.

[0040] (2) The present invention adopts a multi-channel attention mechanism to adaptively weight and fuse attribute enhancement graph, semantic enhancement graph and topology enhancement graph, and further adaptively fuses global enhancement heterogeneous graph and high-order connected graph through multi-channel attention mechanism, so that different types of structural information can participate in the construction of optimization recommendation graph according to their contribution, avoiding the accumulation of invalid information caused by fixed weights or simple superposition.

[0041] (3) This invention constructs a high-order connected graph by calculating the node connectivity relationships of different orders, which can capture indirect associations formed through multiple intermediate nodes and combine the local enhancement information in the relational subgraph with the global high-order connectivity information in the heterogeneous graph. This is beneficial for mining potential associations between users and items that are not directly recorded by the original interaction data, and alleviates the problem of sparse interaction data in recommendation scenarios.

[0042] (4) The present invention constructs a joint loss function including recommendation ranking loss, graph structure reconstruction loss and structure regularization term, and performs end-to-end joint optimization of graph structure parameters and graph neural network encoder parameters, so that the optimized graph structure is adapted to the user preference modeling target, while retaining effective structural information and suppressing redundant connections, providing a reliable foundation for learning the embedding representation of user nodes and item nodes and generating the Top-K recommendation list. Attached Figure Description

[0043] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 The flowchart of the personalized recommendation method based on multi-view heterogeneous attribute graph structure learning of the present invention is shown.

[0044] Figure 2 The diagram shows a comparison of the micro-average F1 values ​​of ablation experiments of different modules on the Yelp, DBLP, and ACM datasets based on the present invention.

[0045] Figure 3 The figure shows a comparison of the macro-average F1 scores of ablation experiments of different modules on Yelp, DBLP, and ACM datasets based on the present invention. Detailed Implementation

[0046] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] like Figure 1 The personalized recommendation method shown here, based on multi-view heterogeneous attribute graph structure learning, specifically includes the following steps: S1. Obtain node data, node attribute data, and interaction relationship data between nodes in the recommendation scenario, construct a heterogeneous recommendation attribute graph. The heterogeneous recommendation attribute graph includes multiple node types, multiple edge types, and node attribute features. The multiple node types include at least user nodes and item nodes. Establish a unified node index space according to all nodes in the heterogeneous recommendation attribute graph.

[0048] S2: The heterogeneous recommendation attribute graph is split into multiple relation subgraphs based on the edge type. Attribute enhancement, semantic enhancement, and topological enhancement are performed on each relation subgraph to generate three types of enhanced relation subgraphs.

[0049] S3 adaptively fuses the three types of enhanced relational subgraphs of each relational subgraph through a multi-channel attention mechanism to obtain the local enhanced graph of each relational subgraph. The local enhanced graphs of all relational subgraphs are then mapped to a unified node space and aggregated to obtain the global enhanced graph.

[0050] S4. Based on the adjacency matrix of the heterogeneous recommendation attribute graph in the unified node index space, calculate the node connectivity from first order to the preset highest order, and fuse the node connectivity of different orders to obtain a high-order connected graph.

[0051] S5 uses a multi-channel attention layer to adaptively fuse the global enhancement graph and the higher-order connected graph to generate the final optimized recommendation graph.

[0052] S6 optimizes the input graph neural network encoder of the recommendation graph, aggregating neighborhood information to generate low-dimensional embedded representations of user nodes and item nodes.

[0053] S7. Construct a joint loss function that includes recommendation ranking loss, graph structure reconstruction loss and structure regularization term, and backpropagate to iteratively optimize the recommendation graph parameters and graph neural network encoder parameters.

[0054] S8 calculates the target user's preference score for all candidate items based on the trained user and item embeddings, and generates a Top-K recommendation list after sorting.

[0055] Specifically, the heterogeneous recommendation attribute graph in step S1 The definition is as follows: ; in, This represents a set of nodes, including: user nodes, merchant nodes, service category nodes, and rating level nodes; The set of edges includes: rating edges between users and merchants, friend edges between users, subordinate edges between merchants and service categories, and belonging edges between merchants and rating levels. This represents the set of types in the graph, including the set of node types. and edge type set ; : This is a node type mapping function used to determine the node type to which any given node belongs; : This is an edge type mapping function used to determine the relation type to which any connecting edge belongs; The node attribute matrix includes user nodes, whose attribute features may include user activity, historical average score, and user behavior statistics; and merchant nodes, whose attribute features may include geolocation code, category tag features, average star rating, number of reviews, and price range.

[0056] Heterogeneous recommendation attribute graphs include: user and merchant rating subgraphs. User-to-user social subgraph Merchant and service category sub-diagram Merchant and rating level sub-graph .

[0057] User and Merchant Rating Subgraph This includes all user nodes and merchant nodes, with edges representing interactive connections formed by users rating merchants. The corresponding adjacency matrix is ​​as follows: . It is the core interactive carrier of recommendation tasks, carrying the user's preference signal for merchants, and can be accompanied by a rating of one to five stars to represent the strength of the preference.

[0058] User-to-user social subgraph : Includes only user nodes, edges represent friend relationships between users, and the corresponding adjacency matrix is , Used to supplement social collaboration information and uncover potential connections between like-minded users.

[0059] Merchant and service category sub-diagram This includes all merchant nodes and service category nodes. Edges represent the membership relationships of merchants within a specific service category. The corresponding adjacency matrix is ​​as follows: , Used to supplement category attribute information and depict the category similarity between merchants.

[0060] Merchant and rating level sub-map This includes all merchant nodes and rating level nodes. Edges represent the affiliation of a merchant to a specific star rating level, and the corresponding adjacency matrix is... , Used to supplement reputation attribute information and depict the similarity of reputation levels among merchants.

[0061] Specifically, step S2 involves performing attribute enhancement on each relation subgraph, including: Based on the edge connection type, the heterogeneous recommendation attribute graph is split into multiple relation subgraphs, resulting in a set of relation subgraphs. : ; in, Indicates the first A relational subgraph; The total number of relation subgraphs; the purpose of partitioning relation subgraphs is to preserve the independent semantics of different connection relationships, so that subsequent attribute enhancement, semantic enhancement and topology enhancement can be applied to local structures with the same or similar semantics respectively, avoiding the direct mixing of different relation types and the resulting semantic noise diffusion. Attribute Augmentation Graph: Attribute augmentation graph is constructed based on the similarity of node attributes. It is used to supplement the missing attribute associations in the original topology and at the same time weaken the negative impact of noisy edges. Candidate attribute graphs are generated from three paths: direct attribute calculation, topological constraint propagation, and latent space mapping. Finally, the attribute augmentation graph is obtained through attention fusion.

[0062] A K-nearest neighbor attribute similarity graph is constructed based on the original node attributes, and the node with the most similar attributes is selected for each node. For each neighbor, a sparse attribute similarity graph is constructed, and the calculation formula is: ; in, The preset number of nearest neighbors; For the first The K-nearest neighbor attribute similarity adjacency matrix corresponding to each relation subgraph; Relational subgraph The attribute matrix of the nodes included in it; For the K-nearest neighbor graph construction operation, calculate the attribute similarity between nodes based on the attribute matrix, and retain the node with the highest similarity for each node. Each node.

[0063] This method completes connections entirely based on attribute similarity, without relying on the original topology, and can effectively correct the disconnection problem caused by missing data collection in the original graph.

[0064] To avoid excessive cross-semantic connections generated by pure attribute similarity, a feature propagation attribute graph is constructed by propagating attributes based on the original topology. Attribute similarity is calculated under topological constraints, and multi-head weighted cosine similarity is used to calculate the attribute similarity between nodes. The calculation formula is: ; in, For the number of attention heads; For the first The corresponding dimension weight vector is used to differentiate the importance of different attribute dimensions; This is the Hadamard product (element-by-element product). The cosine similarity function is used. For nodes The original attribute feature vector; For nodes The original attribute feature vector.

[0065] A threshold is set on the similarity results for sparsity filtering, resulting in the feature similarity matrix of the head nodes of the relation subgraph. : ; in, A preset similarity threshold is used to filter out similar links with low confidence. and All are relational subgraphs The node indices in the table represent the two nodes that participate in the attribute similarity calculation.

[0066] Combining the subgraph adjacency matrix with forward and backward feature propagation yields two feature propagation graphs: ; in, For the first The head-end feature propagation graph corresponding to each relational subgraph; For the first Tail feature propagation graphs corresponding to each relational subgraph; Relational subgraph The adjacency matrix; and These are the feature similarity matrices for the head and tail nodes of the relation subgraph, respectively. The construction method and Same, but uses a different set of independent trainable feature dimension weight parameters. and similarity threshold .

[0067] The feature propagation graph binds attribute similarity to the original topology, ensuring that the newly added connections satisfy both attribute similarity and local topological distribution, thus having stronger semantic rationality compared to the pure KNN graph.

[0068] To address the issue of inconsistent attribute dimensions and distributions among heterogeneous nodes, a latent space mapping attribute graph is constructed. First, a type-specific mapping layer projects the original attributes onto a unified low-dimensional latent space. Then, similarity is calculated, and a non-linear mapping is applied to the node attributes to obtain the mapped node attributes. : ; in, For nodes The type of node it belongs to; , These are the trainable weight matrix and bias vector for the corresponding node type, respectively; It is the Sigmoid non-linear activation function.

[0069] Node similarity is calculated in the mapped latent space, and the latent space attribute similarity map is obtained after thresholding. : ; in, (·,·) represents the first The multi-weighted cosine similarity function of each relation subgraph in the latent attribute space is calculated in the same way as... same; and They are nodes and nodes The latent attribute representation vector obtained after passing through the node type-specific mapping layer; This is the threshold for potential spatial attribute similarity. When the similarity between nodes is lower than this threshold, the corresponding element is set to 0.

[0070] The attribute similarity map, forward feature propagation map, backward feature propagation map, and latent space attribute similarity map are treated as four channels and adaptively weighted and fused through a multi-channel attention layer to obtain the final attribute enhancement map. : ; in, It is a multi-channel attention fusion function that automatically learns the weight coefficients of four types of candidate images; For the first The attribute augmentation graph corresponding to each relation subgraph.

[0071] Specifically, in step S2, semantic enhancement is performed on each relation subgraph, as follows: For each relation subgraph, a semantic augmentation graph is constructed. The semantic augmentation graph is used to capture the high-level semantic relations expressed by candidate meta-paths, semantic paths, or relation sequences in the heterogeneous recommendation attribute graph.

[0072] Preset candidate metapath set : ; in, For the first There are 10 candidate meta-paths; each candidate meta-path consists of alternating node and edge types, corresponding to an interpretable semantic relationship.

[0073] The semantic similarity graph of a single candidate metapath is constructed by calculating the semantic similarity between nodes based on the node representations of the candidate metapaths, and then filtering by a threshold to obtain the semantic candidate graph of the corresponding candidate metapath. : ; in, For nodes In the Semantic representation vectors under each candidate metapath; For nodes In the Semantic representation vectors under each candidate metapath; This is a similarity calculation function; This is the semantic similarity threshold.

[0074] The core fusion step of the semantic enhancement branch is used to adaptively merge the semantic similarity maps corresponding to multiple candidate meta-paths into the final semantic enhancement map through an attention mechanism: ; in, For semantic channel attention fusion function; For the first The semantic augmentation graph corresponding to each relation subgraph.

[0075] Specifically, in step S2, topology enhancement is performed on each relational subgraph, as follows: A graph diffusion method is employed to construct a topology-enhanced graph for each relational subgraph. This graph diffuses local topological information from the original connection structure, strengthens potential connections consistent with the original structure, and alleviates the insufficient structural representation caused by edge sparsity. For relational subgraphs… adjacency matrix First, self-loop addition and normalization are performed to obtain a normalized adjacency matrix; then, the topological correlation between nodes is calculated based on the diffusion kernel function, random walk transition matrix, or personalized PageRank. ; ; in, This is the degree diagonal matrix of the correspondence subgraph; It is the identity matrix; Represents the adjacency matrix of the relational subgraph; For the current node; The adjacent nodes corresponding to the current node.

[0076] The topology correlation matrix is ​​thresholded, Top-K truncated, normalized, or sparsified to form a topology enhancement graph: ; in, Indicates the first The topology enhancement graph corresponding to each relational subgraph; The restart probability hyperparameter controls the propagation range of the graph diffusion, and its value range is [value range missing]. .

[0077] Topology-enhanced graphs differ from attribute-enhanced graphs and semantic-enhanced graphs. Attribute-enhanced graphs mainly rely on the similarity of node attributes, semantic-enhanced graphs mainly rely on the semantics of heterogeneous paths, and topology-enhanced graphs mainly rely on the original connection structure and diffusion propagation relationships. The three describe node relationships from different perspectives, which is beneficial for subsequent fusion to obtain more robust local enhanced graphs.

[0078] Specifically, step S3 includes the following steps: Multi-channel attention fusion and global enhancement graph generation: A multi-channel attention mechanism is used to fuse multi-view structural information and generate an optimized recommendation graph.

[0079] S3.1, for each relational subgraph The attribute-enhanced graph, semantic-enhanced graph, and topological-enhanced graph are input into a multi-channel attention layer as different structural channels. The multi-channel attention layer assigns weights based on the contribution of the attribute-enhanced graph, semantic-enhanced graph, and topological-enhanced graph to the current relation subgraph, thus obtaining the local enhancement graph corresponding to the relation subgraph. ; in, Indicates the first The first attention fusion function for each relational subgraph; Indicates the first Local enhancement maps corresponding to each relational subgraph.

[0080] S3.2, Local Enhancement Graph of All Relational Subgraphs The graph is mapped to a unified node space and aggregated. Aggregation methods can include summation, weighted summation, relational attention fusion, normalization fusion, or learnable fusion based on relation importance. To avoid any particular relation type having excessive weight, the edge weights of each relation subgraph can be normalized or relation-level gating can be applied to obtain a globally enhanced graph. : ; in, For global augmentation graph, It includes enhanced multi-view information for all relationship types.

[0081] Specifically, the calculation method for the higher-order connected graph in step S4 is as follows: Higher-order connected graphs are constructed based on first- to c-order connectivity calculations using adjacency matrices. These higher-order connected graphs are used to capture global or semi-global associations between nodes formed through multiple intermediate nodes. ; in, Represents the adjacency matrix; Indicates a second-order connectivity relation; express Order-order connectivity; It can be set according to graph size, average path length, business requirements or validation set performance, and can also be used as a learnable or searchable parameter; For a high-order connected graph, integrating 1st to 2nd order graphs... All nodes of the order are associated.

[0082] To avoid introducing too many dense connections in higher-order connectivity computation, in one alternative implementation, each order matrix can be normalized, truncated, sparsified, or its order weights adjusted to assign weights to different orders, thus constructing the following weighted higher-order connectivity graph. : ; in, , to This represents the weights corresponding to different orders. Alternatively, only the nodes with the highest higher-order connectivity scores for each node can be retained, thereby reducing computational and storage overhead. Higher-order connected graphs and locally enhanced graphs of relational subgraphs are complementary. Locally enhanced graphs emphasize attribute, semantic, and topological enhancements under specific relation types, while higher-order connected graphs emphasize global connectivity across multi-hop paths. The fusion of the two can simultaneously preserve reliable local structures and potential global associations.

[0083] Specifically, the calculation method for optimizing the recommendation graph in step S5 is as follows: The global augmentation graph and the higher-order connected graph are adaptively weighted and merged to output the final optimized recommendation graph used in the GNN encoder. : ; in, This represents the second attention fusion function, which automatically learns the weights of the local structure and the global higher-order structure. The corresponding adjacency matrix is ​​denoted as Its elements can represent the optimized connection weights between nodes. Through a two-level attention fusion architecture, this method can dynamically adjust the contribution ratio of structural information at different views and scales according to the characteristics of the dataset and the downstream task objectives, and output the optimized topology structure best suited to the task. Attention fusion functions can employ channel attention, relational attention, gating fusion, softmax weight allocation, learned scalar weights, or multilayer perceptron weight generation. During training, attention weights are updated based on recommendation ranking loss and graph structure reconstruction loss, ensuring that graph structure learning aligns with the downstream task objectives.

[0084] Specifically, step S6 includes the following steps: S6.1, normalize the adjacency matrix of the optimized recommendation graph: ; in, for The corresponding degree diagonal matrix; This is the normalized adjacency matrix.

[0085] S6.2 uses a graph convolutional network encoder for encoding. The encoding formula of the two-layer graph convolutional network encoder (GCN) is as follows: ; in, This is the first layer of trainable weight matrix; For input attribute dimensions; For the hidden layer dimension; This is the trainable weight matrix for the second layer; Embed dimensions for preset nodes; It is a non-linear activation function; This is the final output node representation matrix, where each row corresponds to a node's representation vector. Based on the node type and uniform node index, from... User node embeddings and item node embeddings are extracted and used for subsequent recommendation ranking loss calculation and candidate item preference score calculation.

[0086] Specifically, the joint loss function is calculated in step S7 as follows: To address the core task of "ranking first" in personalized recommendations, this implementation employs pairwise ranking loss to optimize the order of user preferences for products, allowing graph structure learning to directly contribute to improving recommendation ranking accuracy. The recommendation ranking supervision loss utilizes Bayesian Personalized Ranking (BPR) loss. The specific definition is as follows: ; in, Optimize the recommendation graph for the target user node. User nodes in; For positive sample products, i.e., users Product nodes that have already generated interactions; Negative sample products, i.e., users Product nodes that have not generated any interaction are randomly sampled from the global candidate product pool; To train the set of triples, each triple... Indicates: User For positive sample products The preference level for the negative sample items was higher than that for the negative sample items. ; To optimize the recommended image User obtained through encoding The embedding vector; , Positive sample products Negative sample products The embedding vector; The Sigmoid non-linear activation function is used to map the preference score difference to the (0,1) interval, representing the probability of a correct ranking; This is the set of all trainable weight matrices for a graph neural network encoder. The weight decay coefficient is used to apply L2 regularization to the parameters of the graph neural network encoder to prevent overfitting. is the square of the Frobenius norm of the matrix; This is a transpose operation.

[0087] It directly optimizes the relative preference order of users for products, which is more in line with the business goals of the recommendation system; through backpropagation, the ranking error is transmitted to the graph structure modeling module and the model is driven to automatically weaken noise edges such as accidental clicks and strengthen high-confidence interaction edges, thus realizing the deep coupling between graph structure and recommendation task.

[0088] The recommendation graph structure is optimized by reconstructing the learned node embeddings in reverse, forcing the node embeddings to retain global topological information. This provides unsupervised learning signals for unlabeled or low-interaction nodes, avoiding overfitting of local ranking supervision signals and improving generalization ability in cold-start and long-tail scenarios. In one implementation, the recommendation graph structure is optimized by reconstructing the node embeddings to retain topological relationships between nodes. For heterogeneous graphs in recommendation scenarios, the graph structure reconstruction loss... The specific form is as follows: ; in, This represents a set of nodes, including: user nodes, merchant nodes, service category nodes, and rating level nodes; For nodes Optimize the recommendation image The set of direct neighbor nodes in the; For nodes The set of negative sampling nodes, from the graph and Nodes without direct connections are randomly selected to construct negative instance constraints; For indicator functions, when node With nodes The value is 1 if the nodes are the same, otherwise the value is 0. For nodes Optimize the recommendation image The embedding vector obtained from learning; Candidate nodes Optimize the recommendation image The embedding vector obtained from learning.

[0089] Applying L1 norm constraints to the optimized graph adjacency matrix controls the sparsity of the graph structure, prevents the model from adding excessive redundant edges, and suppresses low-confidence noise edges, ensuring the interpretability and computational efficiency of the graph structure, while also regularizing the loss. The specific formula is as follows: ; in, For a set of nodes The total number of nodes included in it; The adjacency matrix is ​​the first... Line number The element values ​​of the column represent the nodes. With nodes Connection weights between them; The L1 norm of a matrix is ​​the sum of the absolute values ​​of all its elements. This is the scalar absolute value operator.

[0090] Joint loss function for: ; in, For graph structure reconstruction loss The weighting coefficients are used to adjust the contribution of the graph structure reconstruction objective to the joint loss; For structural regularization loss The weighting coefficients are used to adjust the contribution of graph structure sparse constraints to the joint loss; During training, multi-view structure enhancement, higher-order connectivity modeling, attention fusion, graph neural network encoding, and loss backpropagation can be performed iteratively. After training, the optimized recommendation graph is output. And node representation matrix The optimized recommendation graph structure is used to explain the potential relationships between nodes and serves as input for subsequent product recommendation systems.

[0091] Taking personalized recommendation as an example, this paper explains the process of generating personalized recommendation results based on an optimized recommendation graph structure. In user recommendation scenarios, users, products, merchants, categories, tags, and rating levels can be modeled as nodes in a heterogeneous recommendation attribute graph. User node attributes can include age group, region, interest tags, historical click statistics, purchase preferences, and conversation behavior characteristics; product node attributes can include title text, image features, price range, brand, sales volume, category, and content embedding; merchant node attributes can include merchant level, business category, service rating, and region; category and tag nodes can include name, hierarchy, and semantic embedding; rating level nodes can represent one to five stars or other rating ranges.

[0092] The edges in a heterogeneous recommendation attribute graph can include relationships such as user clicks on products, user favorites of products, user purchases of products, user rating of products, users following other users, users browsing merchants, products belonging to a category, products having tags, products being provided by merchants, merchants belonging to the same industry, and products being similar to other products. Each relationship can be considered as an edge type, and multiple relationship subgraphs can be created based on edge type.

[0093] For each relationship subgraph, attribute augmentation graphs, semantic augmentation graphs, and topology augmentation graphs are constructed separately. For example, for the user-click-product relationship subgraph, an attribute augmentation graph can be constructed based on the similarity between user behavior attributes and product attributes; a semantic augmentation graph can be constructed based on semantic paths such as "user—product—category—product" or "user—product—tag—product"; and a topology augmentation graph can be constructed based on the diffusion and propagation of click relationships. A similar approach can be used to construct augmentation graphs for purchase relationships, rating relationships, and follow relationships.

[0094] A higher-order connected graph is constructed to capture potential interest associations between users and products formed through multiple intermediate nodes. For example, connections between a user and products of the same category via purchased products, or connections between a user and products purchased by that user via a followed user, can all be represented by a higher-order connected graph. Subsequently, an attention mechanism is used to fuse the local augmentation graph and the higher-order connected graph to obtain an optimized recommendation graph structure.

[0095] The optimized recommendation graph structure is input into the graph neural network encoder to obtain user node representations and product node representations. The recommendation score can be calculated in any of the following ways.

[0096] During the training phase, user history such as clicks, purchases, favorites, and ratings can be used as supervision signals, and jointly optimized using graph structure reconstruction loss and regularization loss. For implicit feedback recommendations, positive and negative sample construction and ranking losses can be used; for explicit rating recommendations, rating regression loss or multi-classification loss can be used. This invention can learn a more reliable recommendation graph structure and user and product representations while preserving the semantics of heterogeneous recommendation relationships.

[0097] To verify the method provided by this invention, the following two experiments were conducted: verification of the learning effect of heterogeneous graph structure and verification of core module ablation.

[0098] Validation of Heterogeneous Graph Structure Learning Effect: This experiment is used to verify the effectiveness of the multi-view structure enhancement, higher-order connectivity modeling and graph structure optimization mechanisms in this invention for node representation learning. Three public heterogeneous recommendation attribute graph datasets, Yelp, DBLP and ACM, are selected as validation datasets. The Yelp dataset includes various node types such as users, merchants, services and rating levels. The DBLP dataset includes node types such as papers, authors and conferences. The ACM dataset includes node types such as papers, authors and topics.

[0099] In the experiment, the heterogeneous recommendation attribute graph was first divided into multiple relational subgraphs according to edge type or connection relationship. Then, attribute enhancement graphs, semantic enhancement graphs, and topological enhancement graphs were constructed for each relational subgraph. Furthermore, a local enhancement graph was generated through a multi-channel attention mechanism, and a high-order connected graph was generated through a high-order connectivity modeling module. These were then fused to obtain an optimized recommendation graph structure, which was input into a graph neural network encoder to obtain node representations. Finally, semi-supervised node classification validation was performed on labeled nodes in a public dataset.

[0100] This experiment uses the micro-average F1 score and the macro-average F1 score as evaluation metrics. The micro-average F1 score measures the overall classification accuracy, while the macro-average F1 score measures the average classification performance across different categories. Comparison methods include graph representation learning or graph structure learning methods such as DeepWalk, GCN, GAT, LDS-GNN, Pro-GNN, Geom-GCN, Metapath2vec, HAN, GTN, and HGSL. The validation results are shown in Table 1. Table 1. Results (%) of each model in semi-supervised node classification

[0101] As shown in Table 1, the basic structure learning method of this invention achieves superior node classification performance on three heterogeneous recommendation attribute graph datasets: Yelp, DBLP, and ACM. The results indicate that joint structure modeling using attribute augmentation graphs, semantic augmentation graphs, topology augmentation graphs, and higher-order connected graphs can improve the problems of noisy and missing connections in heterogeneous recommendation attribute graphs, enabling the graph neural network encoder to obtain more discriminative node representations. The node representation learning capability can be further transferred to user recommendation scenarios to learn the embedding representations of user nodes and item nodes.

[0102] Ablation validation of core modules: This experiment conducted a series of ablation experiments to better demonstrate the effectiveness of each module in the proposed model. These ablation studies were carried out under the node classification task.

[0103] The experiment uses node classification accuracy as the basic evaluation metric. Node classification accuracy reflects the class discrimination ability of the learned node embeddings and can serve as an auxiliary verification of the effectiveness of multi-view graph structure optimization and node representation learning. It also provides a foundation for user node and item node representation learning in recommendation scenarios. The experiment was conducted on three heterogeneous datasets: Yelp, DBLP, and ACM. The micro-average F1 score and macro-average F1 score were used as quantification metrics. The performance changes of each module after deletion and the corresponding recommendation scenario value analysis are as follows: 1. Validation of the attribute enhancement module: Remove attribute enhancement (Attr), construct a comparative variant scheme, and the experimental results are as follows. Figure 2 As shown: On the Yelp, DBLP, and ACM datasets, the variant's micro-mean F1 score decreased by 10.53, 0.75, and 2.11 percentage points compared to the full scheme, respectively. Figure 3 As shown, the macro average F1 score decreased by 14.26, 0.82 and 2.00 percentage points, respectively, with the Yelp dataset showing the most significant performance degradation.

[0104] 2. Validation of the semantic enhancement module: The semantic enhancement module in this invention was removed, and a comparative variant scheme was constructed. Experimental results are as follows: Figure 2As shown: On the Yelp, DBLP, and ACM datasets, the variant's micro-mean F1 score decreased by 1.83, 1.80, and 1.64 percentage points compared to the full scheme, respectively. Figure 3 As shown, the macro average F1 score decreased by 1.48, 1.99 and 1.63 percentage points respectively, and all three datasets showed a stable performance decline; The semantic enhancement module mines deep semantic associations based on candidate meta paths in heterogeneous recommendation attribute graphs. In recommendation scenarios, it can capture multi-hop semantic transmission relationships such as "user-product-category" and "user-product-brand", deepen the granularity of user interest mining, and expand the semantic coverage of recommendations. This module can stably improve the embedded semantic expression capabilities in different types of recommendation scenarios, and is an important support for improving recommendation matching accuracy and content diversity.

[0105] 3. Validation of the topology enhancement module: The topology enhancement module in this invention was removed, and a comparative variant scheme was constructed. The experimental results are as follows: Figure 2 As shown: On the Yelp dataset, the micro-average F1 score of this variant decreased by 3.33 percentage points compared to the full scheme; on the DBLP and ACM datasets, the micro-average F1 score decreased by 0.62 and 0.69 percentage points respectively, with significant differences in performance degradation across different scenarios; The topology enhancement module smooths the original topology structure and suppresses interference from isolated noise edges based on a personalized PageRank graph diffusion method. In recommendation scenarios, this module is mainly used to filter topological noise caused by invalid interactions such as accidental clicks and accidental favorites. In the local life recommendation scenario corresponding to the Yelp dataset, the proportion of noisy interactions caused by users accidentally clicking on merchants is higher, so the noise reduction effect of topology enhancement is more significant. However, in scenarios with higher interaction confidence, such as academic literature, the noise reduction effect is relatively gradual. Experimental results verify the value of the topology enhancement module in improving the robustness of high-noise recommendation scenarios.

[0106] 4. Validation of the effectiveness of higher-order connectivity modules: The higher-order connectivity modules in this invention were removed, and a comparative variant scheme was constructed. The experimental results are as follows: Figure 2 As shown: On the Yelp and DBLP datasets, the variant's micro-mean F1 score decreased by 1.69 and 1.04 percentage points compared to the full scheme, respectively; on the ACM dataset, it decreased by 0.14 percentage points. The performance impact varies depending on the characteristics of the dataset structure. The higher-order connectivity module captures multi-hop potential associations through multi-order power operations on the adjacency matrix. In recommendation scenarios, it can mine indirect interest paths such as "user-product-similar user-product", expand the scope of mining potential user interests, and alleviate the problem of insufficient recommendation coverage caused by sparse interaction data. For recommendation scenarios with long user interaction links and rich social relationships, the gains of higher-order connectivity are more obvious, and it has the potential to improve the long-tail coverage of extended recommendations.

[0107] 5. Validation of the graph structure reconstruction module: Remove the graph structure reconstruction loss term from this invention, construct a comparative variant scheme, and the experimental results are as follows. Figure 2 As shown: On the Yelp, DBLP, and ACM datasets, the variant's micro-mean F1 score decreased by 0.62, 0.19, and 0.46 percentage points compared to the full scheme, respectively. Figure 3 As shown, the macro average F1 score decreased by 0.55, 0.28 and 0.41 percentage points, respectively, and all datasets showed a stable, small performance decline.

[0108] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A personalized recommendation method based on multi-view heterogeneous attribute graph structure learning, characterized in that, Specifically, the steps include the following: S1. Obtain node data, node attribute data, and interaction relationship data between nodes in the recommendation scenario, construct a heterogeneous recommendation attribute graph, and establish a unified node index space according to all nodes in the heterogeneous recommendation attribute graph. S2, split the heterogeneous recommendation attribute graph into multiple relation subgraphs based on the edge type, and perform attribute enhancement, semantic enhancement and topological enhancement on each relation subgraph to generate three types of enhanced relation subgraphs; The attribute enhancement specifically involves: splitting the heterogeneous recommendation attribute graph into multiple relation subgraphs based on the edge connection relationship type to obtain a set of relation subgraphs; constructing a sparse attribute similarity graph; and using multi-head weighted cosine similarity to calculate the attribute similarity between nodes. A threshold is set on the similarity results for sparsification filtering to obtain the feature similarity matrix of the head nodes of the relation subgraph. Forward and backward feature propagation is performed using the subgraph adjacency matrix to obtain two feature propagation maps: the forward feature propagation map and the backward feature propagation map. Node attributes are non-linearly mapped to obtain the mapped node attributes. Node similarity is calculated in the mapped latent space and filtered by a threshold to obtain the latent space attribute similarity map. The attribute similarity map, forward feature propagation map, backward feature propagation map, and latent space attribute similarity map are used as four channels and adaptively weighted and fused through a multi-channel attention layer to obtain the final attribute enhancement map. S3 adaptively fuses the three types of enhanced relation subgraphs of each relation subgraph through a multi-channel attention mechanism to obtain the local enhanced graph of each relation subgraph. The local enhanced graphs of all relation subgraphs are mapped to a unified node space and aggregated to obtain the global enhanced graph. S4. Based on the adjacency matrix of the heterogeneous recommendation attribute graph in the unified node index space, calculate the node connectivity from first order to the preset highest order, and fuse the node connectivity of different orders to obtain a high-order connected graph. S5 uses a multi-channel attention layer to adaptively fuse the global enhancement graph and the high-order connected graph to generate the final optimized recommendation graph; S6 optimizes the input graph neural network encoder of the recommendation graph, aggregating neighborhood information to generate low-dimensional embedded representations of user nodes and item nodes; S7, construct a joint loss function including recommendation ranking loss, graph structure reconstruction loss and structure regularization term, and backpropagate to iteratively optimize the recommendation graph parameters and graph neural network encoder parameters; S8 calculates the target user's preference score for all candidate items based on the trained user and item embeddings, and generates a Top-K recommendation list after sorting.

2. The personalized recommendation method based on multi-view heterogeneous attribute graph structure learning according to claim 1, characterized in that, Heterogeneous Recommendation Attribute Graph in Step S1 The definition is as follows: ; in, This represents a set of nodes, including: user nodes, merchant nodes, service category nodes, and rating level nodes; The set of edges includes: rating edges between users and merchants, friend edges between users, subordinate edges between merchants and service categories, and belonging edges between merchants and rating levels. This represents the set of types in the graph, including the set of node types. and edge type set ; A mapping function for node types; For edge type mapping functions; This is a matrix of node attributes.

3. The personalized recommendation method based on multi-view heterogeneous attribute graph structure learning according to claim 2, characterized in that, Step S2 involves performing attribute enhancement on each relation subgraph, specifically as follows: Based on the connection type of the edges, the heterogeneous recommendation attribute graph is split into multiple relation subgraphs, resulting in a set of relation subgraphs. : ; in, Indicates the first A relational subgraph, The total number of relational subgraphs; Construct a sparse attribute similarity graph, calculated using the following formula: ; in, The preset number of nearest neighbors; For the first The K-nearest neighbor attribute similarity adjacency matrix corresponding to each relation subgraph; Relational subgraph The attribute matrix of the nodes included in it; Graph construction operation for K nearest neighbors; Multi-head weighted cosine similarity is used to calculate the attribute similarity between nodes. The calculation formula is: ; in, For the number of attention heads; For the first The dimension weight vector corresponding to the head; For Hadamah accumulation; The cosine similarity function; For nodes The original attribute feature vector; For nodes The original attribute feature vector; A threshold is set on the similarity results for sparsity filtering, resulting in the feature similarity matrix of the head nodes of the relation subgraph. : ; in, The preset similarity threshold; and All are relational subgraphs Node index in; Combining the subgraph adjacency matrix with forward and backward feature propagation yields two feature propagation graphs: a forward feature propagation graph and a backward feature propagation graph. ; in, For the first Forward feature propagation graphs corresponding to each relation subgraph; For the first Backpropagation graphs corresponding to each relation subgraph; Relational subgraph The adjacency matrix; and These are the feature similarity matrices of the head and tail nodes of the relation subgraph, respectively. Perform a non-linear mapping on the node attributes to obtain the mapped node attributes. : ; in, For nodes The type of node it belongs to; , These are the trainable weight matrix and bias vector for the corresponding node type, respectively; It is the Sigmoid non-linear activation function; Node similarity is calculated in the mapped latent space, and the latent space attribute similarity map is obtained after thresholding. : ; in, For the first The multiple weighted cosine similarity function of each relation subgraph in the latent attribute space; For nodes The latent attribute representation vector obtained after passing through the node type-specific mapping layer; The threshold for the similarity of latent spatial attributes; The attribute similarity map, forward feature propagation map, backward feature propagation map, and latent space attribute similarity map are treated as four channels and adaptively weighted and fused through a multi-channel attention layer to obtain the final attribute enhancement map. : ; in, For multi-channel attention fusion function; For the first The attribute-enhanced graph corresponding to each relational subgraph.

4. The personalized recommendation method based on multi-view heterogeneous attribute graph structure learning according to claim 3, characterized in that, Step S2 involves performing semantic enhancement on each relation subgraph, specifically as follows: Preset candidate metapath set : ; in, For the first Candidate metapaths; The semantic similarity between nodes is calculated based on the candidate metapath node representation, and the semantic candidate graph is obtained after threshold filtering. : ; in, For nodes In the Semantic representation vectors under each candidate metapath; For nodes In the Semantic representation vectors under each candidate metapath; This is a similarity calculation function; This is the semantic similarity threshold; The graph is adaptively merged into a final semantically enhanced graph using an attention mechanism. ; in, For semantic channel attention fusion function; For the first The semantic augmentation graph corresponding to each relation subgraph.

5. The personalized recommendation method based on multi-view heterogeneous attribute graph structure learning according to claim 4, characterized in that, Step S2 involves performing topology enhancement on each relation subgraph, specifically as follows: For relational subgraphs adjacency matrix First, self-loop addition and normalization are performed to obtain a normalized adjacency matrix; then, the topological correlation between nodes is calculated based on the diffusion kernel function, random walk transition matrix, or personalized PageRank. ; ; in, This is the degree diagonal matrix of the correspondence subgraph; It is the identity matrix; Represents the adjacency matrix of the relational subgraph; For the current node; The adjacent nodes corresponding to the current node; The topology relevance matrix is ​​thresholded, Top-K truncated, normalized, or sparsified to form a topology enhancement graph. ; in, Indicates the first The topology enhancement graph corresponding to each relational subgraph; This is the restart probability hyperparameter.

6. The personalized recommendation method based on multi-view heterogeneous attribute graph structure learning according to claim 5, characterized in that, Step S3 specifically includes the following steps: S3.1, for each relational subgraph The attribute-enhanced graph, semantic-enhanced graph, and topological-enhanced graph are input into a multi-channel attention layer as different structural channels. The multi-channel attention layer assigns weights based on the contribution of the attribute-enhanced graph, semantic-enhanced graph, and topological-enhanced graph to the current relation subgraph, thus obtaining the local enhancement graph corresponding to the relation subgraph. ; in, Indicates the first The first attention fusion function for each relational subgraph; Indicates the first Local augmentation maps corresponding to each relational subgraph; S3.2, Local Enhancement Graph of All Relational Subgraphs Mapping to a unified node space and aggregating them yields a global augmented graph. : ; in, This is a global enhancement graph.

7. The personalized recommendation method based on multi-view heterogeneous attribute graph structure learning according to claim 6, characterized in that, The specific calculation method for the higher-order connected graph in step S4 is as follows: Constructing high-order connected graphs based on first- to c-order connectivity calculations using adjacency matrices: ; in, Represents the adjacency matrix; Indicates a second-order connectivity relation; express Order-order connectivity; It is a high-order connected graph; Normalize, truncate, sparsify, or adjust the order weights of each matrix, and set weights for different orders to construct the following weighted high-order connected graph. : ; in, , to This represents the weights corresponding to different orders.

8. The personalized recommendation method based on multi-view heterogeneous attribute graph structure learning according to claim 7, characterized in that, The specific calculation method for optimizing the recommendation graph in step S5 is as follows: The global augmentation graph and the higher-order connected graph are adaptively weighted and merged to output the final optimized recommendation graph used in the GNN encoder. : ; in, This represents the second attention fusion function; The corresponding adjacency matrix is ​​denoted as .

9. The personalized recommendation method based on multi-view heterogeneous attribute graph structure learning according to claim 1, characterized in that, Step S6 specifically includes the following steps: S6.1, normalize the adjacency matrix of the optimized recommendation graph: ; in, To optimize the adjacency matrix of the recommendation graph The corresponding degree diagonal matrix; This is the normalized adjacency matrix; S6.2 uses a graph convolutional network encoder for encoding. The encoding formula of the two-layer graph convolutional network encoder (GCN) is as follows: ; in, This is the first layer of trainable weight matrix; For input attribute dimensions; For the hidden layer dimension; This is the trainable weight matrix for the second layer; Embed dimensions for preset nodes; It is a non-linear activation function; This is the final output matrix representing the nodes.

10. A personalized recommendation method based on multi-view heterogeneous attribute graph structure learning according to claim 1, characterized in that, The specific calculation method of the joint loss function in step S7 is as follows: The recommendation ranking supervision loss adopts Bayesian personalized ranking BPR loss. The specific definition is as follows: ; in, For target user nodes; These are positive sample products; These are negative sample products; To train the set of triples; To optimize the recommended image User obtained through encoding The embedding vector; and Positive sample products and negative sample products The embedding vector; It is the Sigmoid non-linear activation function; This is the set of all trainable weight matrices for a graph neural network encoder. This is the weight decay coefficient; is the square of the Frobenius norm of the matrix; This is a transpose operation; It is a logarithmic function; Graph structure reconstruction loss The specific form is as follows: ; in, For nodes Optimize the recommendation image The set of direct neighbor nodes in the; For nodes The set of negative sampling nodes; For indicator functions; For nodes Optimize the recommendation image The embedding vector obtained from learning; Candidate nodes Optimize the recommendation image The embedding vector obtained from learning; Regularization loss The specific formula is as follows: ; in, For a set of nodes The total number of nodes included in it; The first in the adjacency matrix Line number The element values ​​of the column; Let L1 be the norm of the matrix; The scalar absolute value operator; Joint loss function for: ; in, For graph structure reconstruction loss Weighting coefficients; For structural regularization loss The weighting coefficients.

Citation Information

Patent Citations

  • Learable enhanced comparison recommendation method capable of adaptively fusing multiple views

    CN120723974A

  • Recommendation method and system based on large language model enhancement and structured alignment

    CN121188292A