Spectral graph convolution network heterogeneous graph representation learning method based on path collaborative graph enhancement
By constructing a path collaboration graph and combining it with a graph attention network, and employing a collaborative multinomial spectral filter and a learnable positive definite diagonal matrix, the problems of insufficient path importance differentiation and missing semantic collaboration in heterogeneous graph representation learning are solved, thereby improving the accuracy and robustness of node classification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUIZHOU NORMAL UNIVERSITY
- Filing Date
- 2026-03-09
- Publication Date
- 2026-06-09
AI Technical Summary
Existing heterogeneous graph representation learning methods lack fine-grained distinctions between path importance and collaborative interactions between meta-paths, leading to information redundancy or underutilization.
By constructing a path collaboration graph and combining it with a graph attention network to learn path interaction weights, and employing a collaborative multinomial spectral filter and a learnable positive definite diagonal matrix, fine-grained importance distinction and semantic collaborative capture of metapaths are achieved.
It improves the node classification performance of heterogeneous graphs, enhances the ability to perceive complex heterogeneous semantics and robustness, and improves the accuracy and stability of downstream tasks.
Smart Images
Figure CN121787494B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of graph neural network technology, specifically to a heterogeneous graph representation learning method for spectral graph convolutional networks based on path-cooperative graph enhancement. Background Technology
[0002] Graph-structured data is widely present in various real-world scenarios, such as social networks, academic citation networks, and e-commerce transaction networks. Among them, heterogeneous graphs, due to their diverse types of nodes and edges, can carry richer semantic information than homogeneous graphs, making them an important research direction for graph neural networks. The core objective of heterogeneous graph representation learning is to map the nodes in a heterogeneous graph to a low-dimensional, dense vector space while preserving the graph's structural information and semantic features, providing strong support for subsequent downstream tasks such as node classification, link prediction, and community detection.
[0003] Existing heterogeneous graph representation learning methods largely rely on manual selection of meta-paths. However, this method lacks a solid theoretical basis and heavily depends on the experience of domain experts, making it difficult to scale to diverse application scenarios when dealing with large-scale, complex heterogeneous graphs. To address this issue, spectral heterogeneous graph convolutional networks (such as PSHGCN) have recently emerged. This method uses multinomial filters to approximate spectral convolution, achieving a certain degree of automated semantic learning. However, this technique also has some drawbacks:
[0004] 1. Lack of fine-grained path importance learning: Although existing methods (such as PSHGCN) learn weights of different orders, they treat different meta-paths of the same order equally and ignore fine-grained semantic differences.
[0005] 2. Lack of collaborative interaction between meta-paths: Existing methods (such as H2SGNN) usually process each meta-path subgraph independently, ignoring the semantic overlap or complementary relationships between different meta-paths (such as "author-paper-author" and "author-institution-author"), resulting in information redundancy or underutilization.
[0006] Based on this, we now present a heterogeneous graph representation learning method for spectral graph convolutional networks based on path collaborative graph enhancement, which can eliminate the drawbacks of existing technical solutions. Summary of the Invention
[0007] The purpose of this invention is to provide a heterogeneous graph representation learning method for spectral graph convolutional networks based on path collaboration graph enhancement, so as to solve the problem that existing methods in the background art cannot distinguish the importance of meta-paths in a fine-grained manner and capture the collaboration relationships between paths.
[0008] To achieve the above objectives, the present invention provides the following technical solution:
[0009] A heterogeneous graph representation learning method for spectral graph convolutional networks based on path collaborative graph enhancement specifically includes the following steps:
[0010] Step S1: Obtain the heterogeneous graph data to be processed, and use type-aware linear transformation to project the features of different types of nodes onto a unified latent feature space.
[0011] Step S2: Calculate the prior weights of the structure and initialize the importance of the meta-paths based on their order, so that the weights of lower-order meta-paths are higher than those of higher-order meta-paths.
[0012] Step S3: Construct a corresponding path collaboration graph for each set of meta-paths, use the meta-paths as nodes in the path collaboration graph, calculate the semantic similarity between meta-paths as edge attributes, and learn the path interaction weight of each meta-path through a graph attention network.
[0013] Step S4: Combine the structural prior weights and path interaction weights to construct a collaborative multinomial spectral filter for spectral graph convolution;
[0014] Step S5: Perform path co-graph enhancement spectral convolution, introduce a learnable positive definite diagonal matrix into the co-polynomial spectral filter, and perform weighted sum feature transformation on the filtered node features to obtain the final node representation.
[0015] Furthermore, the specific operations of node feature projection in step S1 include:
[0016] For heterogeneous graphs The original features of different node types are projected onto a multilayer perceptron corresponding to the node type. The public space, after projection, has the following characteristics. .
[0017] Furthermore, the formula for calculating the structural prior weights in step S2 is as follows:
[0018] ;
[0019] in, For the first Structural prior weights of order path Let be the order of the metapath. This refers to the hyperparameters used to control the decay rate.
[0020] Furthermore, the specific operations for constructing the path collaboration graph in step S3 include:
[0021] For a given The set of meta-paths treats each meta-path as a node in the path collaboration graph;
[0022] Calculate any two metapaths and Three similarity metrics between them;
[0023] If the sum of the three similarity indices exceeds a preset threshold, a connecting edge is established in the path collaboration graph, and the three similarities are concatenated as the attribute vector of the edge.
[0024] Furthermore, the specific operation of learning the path interaction weights in step S3 includes:
[0025] The scalar edge weights are obtained by weighting and summing the attribute vectors of the edges using the learnable attention vectors.
[0026] Graph attention networks are used to perform message passing operations on path collaboration graphs to update the embedding representations of meta-path nodes.
[0027] Based on the updated meta-path node embedding, the normalized interaction weight of each meta-path is calculated by the Softmax function after feature transformation through a multilayer perceptron.
[0028] Furthermore, the cooperative polynomial spectral filter in step S4 is constructed using an additive fusion strategy, as follows:
[0029] ;
[0030] in, This is the representation of a cooperative polynomial spectral filter. and For global control weights, For structural prior weights, For path interaction weights, The normalized adjacency matrix, This is the preset highest order of the metapath. The order of the metapath being traversed is [number]. For the first Rank path set A specific meta-path in the process.
[0031] Furthermore, the specific operation of introducing a learnable positive definite diagonal matrix in step S5 includes: constructing a diagonal matrix. ,in, These are trainable parameters that correspond one-to-one with each node. , Construct a matrix function. The total number of nodes in the heterogeneous graph data is represented by the positive definite diagonal matrix, which is used to ensure the positive definiteness of the cooperative polynomial spectral filter and adaptively adjust the importance weights of different nodes in the convolution process.
[0032] Furthermore, the output calculation formula for the spectral graph convolution of the path co-graph enhancement in step S5 is as follows:
[0033] ;
[0034] in, For the enhanced node representation, It is a positive definite diagonal matrix. These are the node features after projection. This is the transpose symbol.
[0035] Furthermore, the three similarity metrics include Jaccard similarity, LCS similarity, and positional matching similarity. Jaccard similarity is used to capture the set overlap rate of node types, LCS similarity is used to capture sequence consistency, and positional matching similarity is used to capture strict local alignment.
[0036] Furthermore, a heterogeneous graph representation learning system for spectral graph convolutional networks based on path cooperative graph enhancement is also provided, the system comprising:
[0037] The feature projection module is used to project different types of node features in heterogeneous graph data onto a feature space of uniform dimension.
[0038] The weight calculation module is used to calculate the structural prior weights based on order decay and the path interaction weights based on the path collaboration graph learning path.
[0039] The filtering module is used to construct a collaborative multinomial spectral filter that combines structural prior weights and path interaction weights.
[0040] The convolution module performs spectral graph convolution operations that enhance the path co-graph with a learnable positive definite diagonal matrix, generating the final node representation.
[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0042] This invention constructs a path collaboration graph and combines it with a graph attention network to learn path interaction weights, achieving fine-grained differentiation of meta-path importance. This accurately captures key semantic information and effectively improves the perception of complex heterogeneous semantics. Furthermore, this invention proposes a novel filter form, the collaborative multinomial spectral filter, which employs an additive strategy to fuse structural prior weights and path interaction weights, enhancing the filter's robustness and expressive power. In addition, this invention introduces a learnable positive definite diagonal matrix into the spectral convolution, effectively reflecting node-level importance and ensuring the well-posedness of the spectral optimization problem. Through these operations, this invention demonstrates high accuracy and robustness in downstream tasks such as node classification, showing promising application prospects. Attached Figure Description
[0043] Figure 1 This is a schematic diagram of the method steps of the present invention;
[0044] Figure 2 This is a schematic diagram illustrating the construction of the path collaboration graph of the present invention;
[0045] Figure 3 This is a schematic diagram of the cooperative polynomial spectral filter and convolution calculation of the present invention;
[0046] Figure 4 This is a schematic diagram of the system structure of the present invention;
[0047] Figure 5 This is a schematic diagram of the structure of the electronic device of the present invention;
[0048] Figure labeling: Feature projection module 10, weight calculation module 20, filtering module 30, convolution module 40. Detailed Implementation
[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0050] Example 1
[0051] In this embodiment, as Figures 1-3 As shown, the heterogeneous graph representation learning method for spectral graph convolutional networks based on path collaborative graph enhancement specifically includes the following steps:
[0052] Step S1: Obtain the heterogeneous graph data to be processed, and use type-aware linear transformation to project the features of different types of nodes onto a unified latent feature space. Heterogeneous graphs refer to graph structure data containing two or more different types of nodes and edges, which can carry richer semantic information than homogeneous graphs. For example, academic networks contain different types of nodes such as authors, papers, and institutions, as well as different types of edges such as author-paper and paper-institution.
[0053] Step S2: Calculate the prior weights of the structure. Initialize the importance of the meta-paths based on their order, making the weights of lower-order meta-paths higher than those of higher-order meta-paths. A meta-path is a path pattern with specific semantics defined in a heterogeneous graph, used to describe the connection relationship between different types of nodes. It is represented as: node type 1 - node type 2 - ... - node type n, such as: author - paper - author, author - institution - author.
[0054] Step S3: Construct a corresponding path collaboration graph for each order of meta-path set, use the meta-paths as nodes in the path collaboration graph, calculate the semantic similarity between meta-paths as edge attributes, and learn the path interaction weight of each meta-path through a graph attention network. The path collaboration graph is an auxiliary graph constructed to model the semantic relationship between different meta-paths of the same order.
[0055] Step S4: Combining structural prior weights and path interaction weights, construct a collaborative multinomial spectral filter for spectral graph convolution. Spectral graph convolution is a convolution operation implemented based on the spectral decomposition theory of graphs. By designing filters in the spectral domain of the graph to transform node features, the global structural information and semantic features of the graph can be captured.
[0056] Step S5: Perform path co-graph enhancement spectral graph convolution, introduce a learnable positive definite diagonal matrix into the co-polynomial spectral filter, and perform weighted sum feature transformation on the filtered node features to obtain the final node representation.
[0057] In this embodiment, the present invention can learn the importance of meta-paths in a fine-grained manner and capture the semantic interactions between meta-paths through the above steps, which solves the problem that existing spectral heterogeneous graph convolution cannot distinguish the importance of fine-grained paths and lacks semantic cooperation, and effectively improves the classification performance of heterogeneous graph nodes.
[0058] Specifically, node feature projection refers to projecting different types of nodes in a heterogeneous graph onto a shared latent feature space. The node feature projection operation in step S1 includes:
[0059] For heterogeneous graphs ,in, For a set of nodes, Given a set of edges, a multilayer perceptron (MLP) corresponding to the node type is used to project the original features of different node types onto a dimension of 1. The public space, after projection, has the following characteristics. ;
[0060] Specifically, structural prior weights refer to a heuristic mechanism that assigns higher weights to lower-order path elements (such as...). (Form), to suppress noise and semantic ambiguity caused by higher-order metapaths, the formula for calculating the structural prior weights in step S2 is: ,in, For the first Structural prior weights of order path Let be the order of the metapath. To control the hyperparameters of the decay rate, experiments have shown that when Time (i.e., weight is) This can achieve better initialization results;
[0061] Specifically, this invention constructs a path collaboration graph for each order of meta-path set. In the path collaboration graph, nodes represent meta-paths, and edges represent the similarity between meta-paths. Edges are constructed by calculating three similarities: Jaccard, LCS (Longest Common Subsequence), and position matching. A graph attention network is used to perform message passing on the collaboration graph, thereby learning path interaction weights that reflect the collaboration relationship between paths. The operation of constructing the path collaboration graph in step S3 includes:
[0062] For a given The set of meta-paths treats each meta-path as a node in the path collaboration graph;
[0063] Calculate any two metapaths and Three similarity metrics between them;
[0064] If the sum of the three similarity indices exceeds a preset threshold, then a connecting edge is established in the path collaboration graph, and the three similarities are concatenated as the attribute vector of the edge.
[0065] Specifically, the operation of learning path interaction weights in step S3 includes:
[0066] The learnable attention vector is used to perform a weighted summation of the attribute vectors of the edges to obtain scalar edge weights, which are then input into the SimFusionModule for fusion.
[0067] A graph attention network is used to perform message passing operations on the path collaboration graph, updating the embedding representations of meta-path nodes. A GATv2Conv layer is then used to update the embedding representations of meta-path nodes on the path collaboration graph. ;
[0068] Based on the updated meta-path node embeddings, feature transformation is performed using a multilayer perceptron (MLP), followed by calculation of the normalized interaction weights for each meta-path using the Softmax function. ;
[0069] Specifically, unlike the multiplicative combination of existing technologies, the cooperative polynomial spectral filter of this invention is a novel filter form that employs an additive strategy. The fusion of prior structural weights and path interaction weights enhances the robustness and expressiveness of the filter. The cooperative polynomial spectral filter in step S4 is constructed using an additive fusion strategy, as follows:
[0070] ;
[0071] in, This is the representation of a cooperative polynomial spectral filter. and For global control weights, For structural prior weights, For path interaction weights, The normalized adjacency matrix, This is the preset highest order of the metapath. The order of the metapath being traversed is [number]. For the first Rank path set A specific meta-path in the process. It is an identity matrix used to preserve the characteristics of the nodes themselves;
[0072] Specifically, this invention introduces a learnable positive definite diagonal matrix into spectral convolution. This matrix not only reflects the importance at the node level but also guarantees the positive definiteness of the graph filter, ensuring the well-posedness of the spectral optimization problem. The operation of introducing a learnable positive definite diagonal matrix in step S5 includes: constructing a diagonal matrix. ,in, These are trainable parameters that correspond one-to-one with each node. , Construct a matrix function. The total number of nodes in the heterogeneous graph data is represented by a positive definite diagonal matrix, which is used to ensure the positive definiteness of the cooperative polynomial spectral filter and adaptively adjust the importance weights of different nodes in the convolution process.
[0073] Specifically, the output calculation formula for the spectral graph convolution of the path co-graph enhancement in step S5 is as follows:
[0074] ;
[0075] in, For the enhanced node representation, It is a positive definite diagonal matrix. These are the node features after projection. The transpose symbol is used to finally obtain the node representation. The input is fed into the output layer MLP for classification prediction, and the cross-entropy loss function is used for training;
[0076] Specifically, the three similarity metrics include Jaccard similarity, LCS similarity, and positional matching similarity. Jaccard similarity ( ) is used to capture the overlap rate of sets of node types, LCS similarity ( ) is used to capture sequence consistency, position matching similarity ( This is used to capture strict local alignment; if the sum of the three values exceeds a preset threshold, then... Then, the connection edge and the preset threshold can be adjusted according to theoretical basis, experimental data and actual needs;
[0077] In this embodiment, the present invention addresses the problems of insufficient fine-grained path importance differentiation and lack of meta-path collaborative interaction in existing heterogeneous graph representation learning methods through operations such as node feature projection, structural prior weight calculation, path collaboration graph construction and path interaction weight learning, collaborative polynomial spectral filter construction, and spectral convolution enhanced by path collaboration graph. This method achieves fine-grained adaptive learning of meta-path importance by combining path collaboration graph with structural prior weight. The additive fusion strategy enhances the robustness and expressive power of the collaborative polynomial spectral filter, and the learnable positive definite diagonal matrix improves the attention to key nodes and training stability, thereby demonstrating better performance in downstream tasks such as node classification.
[0078] Example 2
[0079] The difference from Example 1 is that, as in Example 2, ... Figure 4 As shown, the present invention also provides a heterogeneous graph representation learning system for spectral graph convolutional networks based on path cooperative graph enhancement. The heterogeneous graph representation learning system for spectral graph convolutional networks includes:
[0080] Feature projection module 10 is used to project different types of node features in heterogeneous graph data onto a feature space of uniform dimension.
[0081] The weight calculation module 20 is used to calculate the structural prior weights based on order decay and the path interaction weights based on the path collaboration graph learning path.
[0082] Filtering module 30 is used to construct a cooperative multinomial spectral filter that combines structural prior weights and path interaction weights.
[0083] Convolution module 40 is used to perform spectral graph convolution operations that enhance the path co-graph with a learnable positive definite diagonal matrix, generating the final node representation.
[0084] Among them, such as Figure 5 As shown, the present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store executable instructions, and the processor is used to read the executable instructions stored in the memory and execute the spectral graph convolutional network heterogeneous graph representation learning method in Embodiment 1 or the spectral graph convolutional network heterogeneous graph representation learning system in Embodiment 2;
[0085] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the heterogeneous graph representation learning method of spectral convolutional networks as in Embodiment 1 or the heterogeneous graph representation learning system of spectral convolutional networks as in Embodiment 2.
[0086] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A heterogeneous graph representation learning method for spectral graph convolutional networks based on path-cooperative graph enhancement, characterized in that, Specifically, the following steps are included: Step S1: Obtain the heterogeneous graph data to be processed, and use type-aware linear transformation to project the features of different types of nodes onto a unified latent feature space. Step S2: Calculate the prior weights of the structure and initialize the importance of the meta-paths based on their order, so that the weights of lower-order meta-paths are higher than those of higher-order meta-paths. Step S3: Construct a corresponding path collaboration graph for each set of meta-paths, use the meta-paths as nodes in the path collaboration graph, calculate the semantic similarity between meta-paths as edge attributes, and learn the path interaction weight of each meta-path through a graph attention network. Step S4: Combine the structural prior weights and path interaction weights to construct a collaborative multinomial spectral filter for spectral graph convolution; Step S5: Perform path co-graph enhancement spectral graph convolution, introduce a learnable positive definite diagonal matrix into the co-polynomial spectral filter, and perform weighted sum feature transformation on the filtered node features to obtain the final node representation. The formula for calculating the structural prior weights in step S2 is as follows: ； in, For the first Structural prior weights of order path Let be the order of the metapath. Hyperparameters for controlling the decay rate; The cooperative polynomial spectral filter in step S4 is constructed using an additive fusion strategy, as follows: ； in, This is the representation of a cooperative polynomial spectral filter. and For global control weights, For structural prior weights, For path interaction weights, The normalized adjacency matrix, This is the preset highest order of the metapath. The order of the metapath being traversed is [number]. For the first Rank path set A specific meta-path in the process.
2. The heterogeneous graph representation learning method for spectral graph convolutional networks based on path cooperative graph enhancement according to claim 1, characterized in that, The specific operations of node feature projection in step S1 include: For heterogeneous graphs The original features of different node types are projected onto a multilayer perceptron corresponding to the node type. The public space, after projection, has the following characteristics. .
3. The heterogeneous graph representation learning method for spectral graph convolutional networks based on path cooperative graph enhancement according to claim 2, characterized in that, The specific operations for constructing the path collaboration graph in step S3 include: For a given The set of meta-paths treats each meta-path as a node in the path collaboration graph; Calculate any two metapaths and Three similarity metrics between them; If the sum of the three similarity indices exceeds a preset threshold, a connecting edge is established in the path collaboration graph, and the three similarities are concatenated as the attribute vector of the edge.
4. The heterogeneous graph representation learning method for spectral graph convolutional networks based on path cooperative graph enhancement according to claim 3, characterized in that, The specific operations for learning path interaction weights in step S3 include: The scalar edge weights are obtained by weighting and summing the attribute vectors of the edges using the learnable attention vectors. Graph attention networks are used to perform message passing operations on path collaboration graphs to update the embedding representations of meta-path nodes. Based on the updated meta-path node embedding, the normalized interaction weight of each meta-path is calculated by the Softmax function after feature transformation through a multilayer perceptron.
5. The heterogeneous graph representation learning method for spectral graph convolutional networks based on path cooperative graph enhancement according to claim 1, characterized in that, The specific operations for introducing a learnable positive definite diagonal matrix in step S5 include: constructing a diagonal matrix. ,in, These are trainable parameters that correspond one-to-one with each node. , Construct a matrix function. The total number of nodes in the heterogeneous graph data is represented by the positive definite diagonal matrix, which is used to ensure the positive definiteness of the cooperative polynomial spectral filter and adaptively adjust the importance weights of different nodes in the convolution process.
6. The heterogeneous graph representation learning method for spectral graph convolutional networks based on path cooperative graph enhancement according to claim 1, characterized in that, The output calculation formula for the spectral convolution of the path co-graph enhancement in step S5 is as follows: ； in, For the enhanced node representation, It is a positive definite diagonal matrix. These are the node features after projection. This is the transpose symbol.
7. The heterogeneous graph representation learning method for spectral graph convolutional networks based on path cooperative graph enhancement according to claim 3, characterized in that, The three similarity metrics include Jaccard similarity, LCS similarity, and positional matching similarity. Jaccard similarity is used to capture the set overlap rate of node types, LCS similarity is used to capture sequence consistency, and positional matching similarity is used to capture strict local alignment.
8. The heterogeneous graph representation learning method for spectral graph convolutional networks based on path cooperative graph enhancement according to claim 1, characterized in that, It also relates to a heterogeneous graph representation learning system for spectral graph convolutional networks based on path cooperative graph enhancement, the spectral graph convolutional network heterogeneous graph representation learning system comprising: The feature projection module (10) is used to project different types of node features in heterogeneous graph data onto a feature space of uniform dimension. The weight calculation module (20) is used to calculate the structural prior weights based on order decay and the path interaction weights based on the path collaboration graph learning path; The filtering module (30) is used to construct a collaborative polynomial spectral filter that combines structural prior weights and path interaction weights; The convolution module (40) performs a spectral convolution operation that enhances the path co-graph with a learnable positive definite diagonal matrix, generating the final node representation.