A Multi-View Graph Neural Network Aggregation Enhancement Method Based on Neighbor Selection

Abstract

This invention discloses a method for enhancing multi-view graph neural network aggregation based on neighbor selection, comprising: evaluating neighbor importance: calculating a similarity score matrix between nodes in the graph based on the adjacency matrix and feature matrix of the graph, used to measure the importance of neighbor nodes; constructing a node filtering model: introducing the concept of retention ratio, i.e., the proportion of important neighbor nodes retained to the total number of neighbor nodes, and constructing a functional relationship model between the retention ratio and the average similarity score of the retained nodes; balancing neighbor importance and the number of neighbor connections: based on the functional relationship model, introducing a balance factor β as a hyperparameter to balance neighbor importance and the number of neighbor connections; information aggregation: designing an enhanced cross-view aggregation framework to realize the mutually reinforcing process of intra-graph aggregation and cross-graph attention aggregation; this invention can more effectively process and analyze multi-view graph structure data.

Description

Technical Field

[0001] This invention relates to the field of graph neural network technology, and in particular to a method for enhancing multi-view graph neural network aggregation based on neighbor selection. Background Technology

[0002] In the fields of deep learning and data science, graph neural networks (GNNs) have become a powerful tool for processing structured data, especially demonstrating unique advantages in handling graph-structured data. Graph data naturally appears in various applications, such as social networks, recommender systems, knowledge graphs, and various types of network systems. With the development of multi-view learning, researchers have begun to focus on how to synthesize information from multiple perspectives to gain a more comprehensive understanding of data structures and content. Multi-view graph neural networks (GNNs) have been developed driven by this need; they combine information from different views to provide richer and more accurate data analysis and predictions than a single view.

[0003] Despite the success of multi-view GNNs in various fields, several challenges remain in effectively extracting and aggregating graph structure features from different views. Existing techniques primarily rely on simple neighbor aggregation strategies and standard attention mechanisms, which can lead to information loss or oversmoothing in some cases. For example, traditional GNN models such as GCN (Graph Convolutional Network) or GAT (Graph Attention Network) mainly focus on feature fusion from local neighbors, neglecting the potential connections and interactions between different views.

[0004] In summary, existing multi-view GNN models typically rely on traditional neighbor aggregation mechanisms, such as simple averaging or weighted summation. These methods may lead to over-smoothing of information, and existing models struggle to capture unique and critical information from different views. Furthermore, existing models often fail to fully utilize the dependencies and complementarities between views when processing multi-view data, which limits the depth and effectiveness of information fusion. In addition, neighbor importance evaluation in existing technologies is mostly conducted in a supervised manner, relying on label information, which has extremely limited applicability in unlabeled data environments. Summary of the Invention

[0005] To address the problems existing in the prior art, the purpose of this invention is to provide a method for enhancing the aggregation of multi-view graph neural networks based on neighbor selection. This invention can more effectively process and analyze multi-view graph structure data.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a multi-view graph neural network aggregation enhancement method based on neighbor selection, comprising the following steps:

[0007] Step 1: Evaluate the importance of neighbors: Based on the adjacency matrix and feature matrix of the graph, calculate the similarity score matrix between nodes in the graph to measure the importance of neighboring nodes;

[0008] Step 2: Construct a node filtering model: Introduce the concept of retention ratio, which is the proportion of important neighbor nodes retained out of the total number of neighbor nodes, and construct a functional relationship model between the retention ratio and the average similarity score of the retained nodes;

[0009] Step 3: Balancing Neighbor Importance and Connection Information: Based on the functional relationship model described above, a balance factor β is introduced as a hyperparameter to weigh neighbor importance and the number of neighbor connections.

[0010] Step 4, Information Aggregation: Design an enhanced cross-view aggregation framework to achieve a mutually reinforcing process of intra-graph aggregation and inter-graph cross-attention aggregation.

[0011] As a further improvement of the present invention, step 1 specifically includes the following steps:

[0012] Step 1.1: Apply Jaccard similarity to the adjacency matrix to measure the similarity between two nodes based on shared connections;

[0013] Step 1.2: Use Cosine similarity on the feature matrix to measure the similarity of features between two nodes;

[0014] Step 1.3: Add the two similarities together to obtain a comprehensive similarity score matrix that considers connection similarity and feature similarity.

[0015] As a further improvement of the present invention, step 2 specifically includes the following steps:

[0016] Step 2.1: Define the retention ratio T as the independent variable and the average similarity score of the retained nodes S as the dependent variable. Assume that the retention ratio and the average similarity score of the retained nodes satisfy the functional relationship S=aT^b+c, where a, b, and c are parameters to be estimated.

[0017] Step 2.2: Randomly sample multiple (T, S) data pairs, and use the curve_fit function of the scipy library to estimate the parameters a, b, and c to obtain the specific function model.

[0018] As a further improvement of the present invention, step 3 specifically includes the following steps:

[0019] Step 3.1: Construct the balance factor constraint βS=(1-β)T, where β∈(0,1); the closer the β value is to 1, the more it emphasizes the importance of retaining neighbor nodes; the closer the β value is to 0, the more it emphasizes retaining sufficient connection information.

[0020] Step 3.2: Substitute the functional relationship model into the constraint conditions of the balance factor to solve for the optimal retention ratio T of each view under the corresponding β value;

[0021] Step 3.3: Based on the optimal retention ratio T, select the top T node pairs in the similarity score matrix, modify the original adjacency matrix, set the elements of the selected node pairs to 1, and set the remaining unselected connection elements to 0, and output the filtered adjacency matrix.

[0022] As a further improvement of the present invention, step 4 specifically includes the following steps:

[0023] Step 4.1: Using the enhanced adjacency matrix, the node features are propagated and aggregated within the graph through the graph convolutional network (GCN) to obtain the aggregated feature tensor within the graph.

[0024] Step 4.2: For the obtained aggregated feature tensor, use an attention network to calculate the cumulative attention scores of other views to the current view, thereby obtaining the feature matrix after inter-graph aggregation;

[0025] Step 4.3: Perform step 4.1 again on the feature matrix after inter-graph aggregation, that is, perform the second intra-graph aggregation and output the feature matrix after the second intra-graph aggregation.

[0026] Step 4.4: Perform final inter-graph aggregation on the feature matrix after secondary aggregation. Convert the feature matrix output in Step 4.3 into vector form and obtain the final feature matrix representation of the multi-view graph through pooling operation, thereby realizing the inter-graph information fusion of the multi-view graph neural network.

[0027] This invention aims to more effectively mine and utilize the structural features and information of multi-view graph data; by introducing a novel node filtering strategy and cross-attention mechanism between views, it improves the depth and breadth of information aggregation, thereby providing more accurate graph representation and prediction performance.

[0028] The beneficial effects of this invention are:

[0029] This invention addresses the specific problems and limitations encountered by existing multi-view graph neural networks (GNNs) when processing complex graphical data, particularly in feature extraction and information aggregation; specifically, it is reflected in:

[0030] 1) Improve the accuracy and depth of information aggregation: This approach aims to improve the over-smoothing problem in the information fusion process of existing multi-view GNNs by using a novel node filtering strategy and cross-attention mechanism, allowing the model to more accurately capture and utilize key information from different views. This method can enhance the model's ability to understand and process complex multi-view data structures.

[0031] 2) Enhance the adaptability and flexibility of the model: By introducing a neighbor selection module based on unsupervised learning, this invention does not rely on traditional labeled data training, thus enhancing the applicability of the model in unlabeled or semi-supervised learning environments. Detailed Implementation

[0032] The embodiments of the present invention will be described in detail below.

[0033] Example

[0034] A method for enhancing multi-view graph neural network aggregation based on neighbor selection is proposed, aiming to more effectively mine and utilize the structural features and information resources of multi-view graph data. The technical solution of this embodiment is described in detail below:

[0035] Multi-view image classification and clustering based on the BikeDC dataset. Data preparation: This embodiment uses the CapitalBikeshare dataset (BikeDC) as the experimental data; specifically, it includes the following steps:

[0036] 1. Neighbor importance assessment:

[0037] For each view's graph data, a similarity score matrix is ​​calculated between nodes in the graph based on its adjacency matrix and feature matrix to measure the importance of neighboring nodes. Specifically:

[0038] a) Use Jaccard similarity on the adjacency matrix to measure the similarity between two nodes based on shared connections;

[0039] b) Use Cosine similarity on the feature matrix to measure the similarity of features between two nodes;

[0040] c) Add the two similarities together to obtain a comprehensive similarity score matrix that considers both connection similarity and feature similarity.

[0041] 2. Node filtering model construction:

[0042] The concept of "retention ratio" is introduced, which is the proportion of important neighbor nodes retained out of the total number of neighbor nodes. A functional relationship model between the retention ratio and the average similarity score of the retained nodes is constructed.

[0043] a) Define the retention ratio T as the independent variable and the average similarity score S of the retained nodes as the dependent variable. Assume that the two satisfy the functional relationship S=aT^b+c, where a, b, and c are parameters to be estimated.

[0044] b) Randomly sample multiple (T, S) data pairs, and use the curve_fit function of the scipy library to estimate parameters a, b, and c to obtain a specific function model.

[0045] 3. Balancing the importance of neighbors with connectivity information:

[0046] Based on the function model obtained in step 2, a balance factor β is introduced as a hyperparameter to balance the importance of neighbors and the number of neighbor connections.

[0047] a) Construct a balance factor constraint βS = (1-β)T, where β ∈ (0, 1). The closer the β value is to 1, the more emphasis is placed on preserving neighboring nodes; the closer the β value is to 0, the more emphasis is placed on preserving sufficient connection information.

[0048] b) Substitute the function model from step 2 into the constraints of the balance factor to solve for the optimal retention ratio T of each view under the corresponding β value;

[0049] c) Based on the optimal retention ratio T obtained in step 3b), select the top T node pairs in the similarity score matrix, modify the original adjacency matrix, set the elements of the adjacency matrix of the selected node pairs to 1, and set the remaining unselected connection elements to 0, and output the filtered adjacency matrix.

[0050] 4. Information Aggregation Module:

[0051] Design an enhanced cross-view aggregation framework to enable mutually reinforcing processes of intra-graph aggregation and inter-graph cross-attention aggregation.

[0052] a) Using the enhanced adjacency matrix from step 3c), the node features are propagated and aggregated within the graph through a graph convolutional network (GCN) to obtain the aggregated feature tensor within the graph.

[0053] b) For the aggregated feature tensor obtained in step 4a), an attention network is used to calculate the cumulative attention scores of other views to the current view. The attention mechanism can adaptively fuse information from different views, overcoming the limitation of treating each view equally, thereby obtaining the feature matrix after graph aggregation;

[0054] c) Using the feature matrix obtained in step 4b), execute step 4a) again, that is, perform the second intra-graph aggregation and output the feature matrix after the second intra-graph aggregation;

[0055] d) Finally, perform graph aggregation. Convert the feature matrix output from step 4c) into vector form, and obtain the final feature matrix representation of the multi-view graph through pooling operations, thus realizing the graph information fusion of the multi-view graph neural network. Testing: The prototype used for comparison is the Graph Attention Network (GAT). Experiments were conducted using the BikeDC dataset, and both classification and clustering tests were performed.

[0056] Dataset: The Capital Bikeshare dataset (BikeDC) is sourced from the Washington, D.C. bicycle system, containing six years of data from 2015 to 2020. Stations are treated as nodes, and traffic flows (edges) are analyzed across three time dimensions: weekdays, weekends, and months. Inbound and outbound flows represent the number of bicycles entering and leaving a station, respectively. Finally, a three-view traffic map extracted within a month is used as an instance, labeled according to the corresponding season (spring, summer, autumn, winter).

[0057] Classification testing: Low-dimensional feature vectors of the graph data are obtained through the method model, and then these feature vectors are input into a Support Vector Machine (SVM) for classification. Classification performance is measured using two metrics: Macro-F1 and Micro-F1, which can comprehensively evaluate the model's performance across various categories in multi-class scenarios.

[0058] Clustering Test: Similarly, low-dimensional feature vectors of the graph data are obtained through the model, and then these feature vectors are clustered using the K-means algorithm. The performance of the feature representations learned by different methods on the clustering task is evaluated using Standardized Mutual Information (NMI) and Adjusted Rand Index (ARI).

[0059] All metrics are better the higher they are. The comparison results are shown in the table below:

[0060] Ma-F1(%) Mi-F1(%) NMI(%) ARI(%) My 60.78 61.76 59.86 37.37 GAT 31.98 48.23 47.41 28.70

[0061] The results show that:

[0062] 1) In classification tasks, the method of this invention significantly outperforms the comparative method GAT in both Macro-F1 and Micro-F1 metrics. Specifically, the Macro-F1 score of this method reaches 60.78%, an improvement of 28.8 percentage points compared to GAT's 31.98%; and in Micro-F1, the method reaches 61.76%, an improvement of 13.53 percentage points compared to GAT's 48.23%. This indicates that the method of this invention can better capture the key features of multi-view graph data and generate more discriminative graph representations, thereby significantly improving classification performance.

[0063] 2) In clustering tasks, the NMI and ARI metrics of the method described in this invention are significantly higher than those of GAT. Specifically, the NMI of this method reaches 59.86%, an improvement of 12.45 percentage points compared to GAT's 47.41%; and the ARI reaches 37.37%, an improvement of 8.67 percentage points compared to GAT's 28.70%. This further verifies the advantages of the method described in this invention in learning the intrinsic structure and similarity of multi-view graph data, demonstrating its effectiveness and adaptability in unsupervised learning scenarios.

[0064] 3) In summary, experiments on the real-world BikeDC dataset demonstrate that the method of this invention achieves significant performance improvements in both classification and clustering tasks, proving its superiority in extracting and aggregating features from multi-view graphs. Specifically, this invention effectively addresses the information oversmoothing problem in existing multi-view GNN models by introducing novel node filtering strategies and cross-attention mechanisms, achieving more accurate and in-depth information fusion. Furthermore, the neighbor selection module based on unsupervised learning enhances the model's applicability and flexibility in unlabeled data environments. These experimental results fully demonstrate that this invention has achieved its intended technical objectives, providing a high-performance and highly adaptable solution for processing complex multi-view graph data.

[0065] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for enhancing multi-view graph neural network aggregation based on neighbor selection, characterized in that, This method is used to classify and cluster multi-view traffic flow data based on bicycle traffic flow datasets. Stations are used as nodes to analyze different traffic flows across three time dimensions: weekdays, weekends, and months. Inbound and outbound traffic flows represent the number of bicycles entering and leaving a station, respectively. Finally, a three-view traffic map extracted within a month is used as an instance and labeled according to the corresponding season. The specific steps include: Step 1: Evaluate the importance of neighbors: Based on the adjacency matrix and feature matrix of the graph, calculate the similarity score matrix between nodes in the graph to measure the importance of neighboring nodes; Step 2: Construct a node filtering model: Introduce the concept of retention ratio, which is the proportion of important neighbor nodes retained out of the total number of neighbor nodes, and construct a functional relationship model between the retention ratio and the average similarity score of the retained nodes; Step 3: Balancing Neighbor Importance and Connection Information: Based on the functional relationship model described above, a balance factor β is introduced as a hyperparameter to weigh neighbor importance and the number of neighbor connections. Step 4, Information Aggregation: Design an enhanced cross-view aggregation framework to achieve a mutually reinforcing process of intra-graph aggregation and inter-graph cross-attention aggregation; Step 4 specifically includes the following steps: Step 4.1: Using the enhanced adjacency matrix, the node features are propagated and aggregated within the graph through the graph convolutional network (GCN) to obtain the aggregated feature tensor within the graph. Step 4.2: For the obtained aggregated feature tensor, use an attention network to calculate the cumulative attention scores of other views to the current view, thereby obtaining the feature matrix after inter-graph aggregation; Step 4.3: Perform step 4.1 again on the feature matrix after inter-graph aggregation, that is, perform the second intra-graph aggregation and output the feature matrix after the second intra-graph aggregation. Step 4.4: Perform final inter-graph aggregation on the feature matrix after secondary aggregation. Convert the feature matrix output in Step 4.3 into vector form and obtain the final feature matrix representation of the multi-view graph through pooling operation, thereby realizing the inter-graph information fusion of the multi-view graph neural network.

2. The multi-view graph neural network aggregation enhancement method based on neighbor selection according to claim 1, characterized in that, Step 1 specifically includes the following steps: Step 1.1: Apply Jaccard similarity to the adjacency matrix to measure the similarity between two nodes based on shared connections; Step 1.2: Use Cosine similarity on the feature matrix to measure the similarity of features between two nodes; Step 1.3: Add the two similarities together to obtain a comprehensive similarity score matrix that considers connection similarity and feature similarity.

3. The multi-view graph neural network aggregation enhancement method based on neighbor selection according to claim 1, characterized in that, Step 2 specifically includes the following steps: Step 2.1: Define the retention ratio T as the independent variable and the average similarity score of the retained nodes S as the dependent variable. Assume that the retention ratio and the average similarity score of the retained nodes satisfy the functional relationship S=aT^b+c, where a, b, and c are parameters to be estimated. Step 2.2: Randomly sample multiple (T, S) data pairs, and use the curve_fit function of the scipy library to estimate the parameters a, b, and c to obtain the specific function model.

4. The multi-view graph neural network aggregation enhancement method based on neighbor selection according to claim 3, characterized in that, Step 3 specifically includes the following steps: Step 3.1: Construct the balance factor constraint βS=(1-β)T, where β∈(0,1); the closer the β value is to 1, the more it emphasizes the importance of retaining neighbor nodes; the closer the β value is to 0, the more it emphasizes retaining sufficient connection information. Step 3.2: Substitute the functional relationship model into the constraint conditions of the balance factor to solve for the optimal retention ratio T of each view under the corresponding β value; Step 3.3: Based on the optimal retention ratio T, select the top T node pairs in the similarity score matrix, modify the original adjacency matrix, set the elements of the selected node pairs to 1, and set the remaining unselected connection elements to 0, and output the filtered adjacency matrix.