Multi-view graph neural network aggregation enhancement method based on neighbor selection

Abstract

The invention discloses a multi-view graph neural network aggregation enhancement method based on neighbor selection, and the method comprises the steps: evaluating the importance of neighbors: calculating a similarity score matrix between nodes in a graph based on an adjacent matrix and a feature matrix of the graph, and enabling the similarity score matrix to be used for measuring the importance of the neighbor nodes; constructing a node filtering model: introducing a retention ratio concept, namely the ratio of the number of reserved important neighbor nodes to the number of total neighbor nodes, and constructing a function relation model of the retention ratio and the average similarity score of the reserved nodes; balancing the neighbor importance and the neighbor connection number: on the basis of the function relation model, introducing a balance factor beta as a hyper-parameter to balance the neighbor importance and the neighbor connection number; information aggregation: designing an enhanced cross-view aggregation framework, and realizing a mutual promotion process of intra-graph aggregation and inter-graph cross attention aggregation; according to the method, the multi-view graph structure data can be processed and analyzed more effectively.

Description

Technical Field

[0001] The present invention relates to the field of graph neural network technology, and in particular to a multi-view graph neural network aggregation enhancement method based on neighbor selection. Background Art

[0002] In the field of deep learning and data science, graph neural networks (GNNs) have become a powerful tool for processing structured data, especially showing their unique advantages in processing graph-structured data. Graph data naturally appears in a variety of applications, such as social networks, recommendation 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 more fully understand the structure and content of data. Multi-view graph neural networks (Multi-view GNNs) were developed under this demand. They combine information from different views to provide richer and more accurate data analysis and predictions than a single view.

[0003] Although multi-view GNNs have achieved certain success in many fields, there are still several challenges in how to effectively extract and aggregate graph structure features from different views. Existing technologies mainly rely on simple neighbor aggregation strategies and standard attention mechanisms, which may lead to information loss or over-smoothing problems in some cases. For example, traditional GNN models such as GCN (Graph Convolutional Network) or GAT (Graph Attention Network) mainly focus on the feature fusion of local neighbors, ignoring the potential correlation and interaction between different views.

[0004] In summary, existing multi-view GNN models usually rely on traditional neighbor aggregation mechanisms such as simple averaging or weighted sum, which may lead to over-smoothing of information, and it is difficult for existing models to capture unique and key information from different views; in addition, existing models often do not fully utilize the dependencies and complementarities between views when processing multi-view data, which limits the depth and effect of information fusion; neighbor importance evaluation in existing technologies is mostly performed in a supervised manner and relies on label information, which has extremely limited applicability in unlabeled data environments. Summary of the invention

[0005] In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a multi-view graph neural network aggregation enhancement method based on neighbor selection, which can more effectively process and analyze multi-view graph structure data.

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

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

[0008] Step 2, constructing a node filtering model: introducing the concept of retention ratio, that is, the ratio of the number 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;

[0009] Step 3: Balance between neighbor importance and connection information: Based on the functional relationship model, a balance factor β is introduced as a hyperparameter to weigh the importance of neighbors and the number of neighbor connections.

[0010] Step 4: Information aggregation: Design an enhanced cross-view aggregation framework to achieve the mutual promotion process of intra-graph aggregation and cross-attention aggregation between graphs.

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

[0012] Step 1.1, use Jaccard similarity on 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 between the features of two nodes;

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

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

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

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

[0018] As a further improvement of the present invention, the step 3 specifically comprises 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 emphasis is placed on retaining neighbor nodes, and the closer the β value is to 0, the more emphasis is placed on retaining sufficient connection information;

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

[0021] Step 3.3: According to the optimal retention ratio T, select the first T node pairs in the similarity score matrix, modify the original adjacency matrix, set the adjacency matrix 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, the step 4 specifically comprises the following steps:

[0023] Step 4.1, using the enhanced adjacency matrix, the graph convolutional network (GCN) is used to propagate and aggregate node features within the graph, and the aggregated feature tensor is obtained;

[0024] Step 4.2: For the obtained aggregated feature tensor, use the attention network to calculate the cumulative attention scores of other views to the current view, so as to obtain the feature matrix after inter-image aggregation;

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

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

[0027] The present 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 a cross-attention mechanism between views to improve the depth and breadth of information aggregation, thereby providing more accurate graph representation and prediction performance.

[0028] The beneficial effects of the present invention are:

[0029] The present invention solves the specific problems and limitations encountered by existing multi-view graph neural networks (GNNs) when processing complex graph data, especially in terms of feature extraction and information aggregation; specifically, it is embodied in:

[0030] 1) Improve the accuracy and depth of information aggregation: A novel node filtering strategy and cross-attention mechanism are used to allow the model to more accurately capture and utilize key information from different views, and improve the over-smoothing problem of existing multi-view GNNs in the information fusion process. This method can improve 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, the present invention does not rely on traditional labeled data training, thereby enhancing the applicability of the model in an unlabeled or semi-supervised learning environment. DETAILED DESCRIPTION

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

[0033] Example

[0034] A multi-view graph neural network aggregation enhancement method based on neighbor selection is intended 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 graph classification and clustering based on BikeDC dataset, data preparation: This embodiment uses CapitalBikeshare dataset (BikeDC) as experimental data; specifically includes the following steps:

[0036] 1. Neighbor importance assessment:

[0037] For each view of the graph data, based on its adjacency matrix and feature matrix, the similarity score matrix between the nodes in the graph is calculated to measure the importance of neighbor 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 between the features of two nodes;

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

[0041] 2. Node filtering model construction:

[0042] The concept of "retention ratio" is introduced, that is, the ratio of the number of important neighbor nodes retained to the total number of neighbor nodes, and 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 the parameters to be estimated.

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

[0045] 3. Balance between neighbor importance and connection information:

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

[0047] a) Construct the 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 retaining neighbor nodes, and the closer the β value is to 0, the more emphasis is placed on retaining sufficient connection information;

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

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

[0050] 4. Information aggregation module:

[0051] An enhanced cross-view aggregation framework is designed to achieve the mutual promotion process of intra-graph aggregation and inter-graph cross-attention aggregation.

[0052] a) Using the enhanced adjacency matrix in step 3c), the graph convolutional network (GCN) is used to propagate and aggregate node features within the graph, and the aggregated feature tensor within the graph is obtained;

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

[0054] c) using the feature matrix obtained in step 4b) to execute step 4a) again, i.e., perform a second intra-graph aggregation, and output the feature matrix after the second intra-graph aggregation;

[0055] d) Finally, perform inter-graph aggregation, convert the feature matrix output in step 4c) into vector form, and obtain the final feature matrix representation of the multi-view graph through pooling operation, so as to realize the inter-graph information fusion of the multi-view graph neural network. Test: The prototype used for comparison is the graph attention network GAT, and experiments are conducted on the BikeDC dataset, and two types of tests are performed: classification and clustering.

[0056] Dataset: Capital Bikeshare dataset (BikeDC) is derived from the bicycle system in Washington, DC, USA, and contains 6 years of data from 2015 to 2020. The stations are taken as nodes to analyze different traffic flows (edges) in three time dimensions: weekdays, weekends, and months. Among them, the inflow and outflow represent the number of bicycles signed in and out of the station, respectively. Finally, the three-view traffic map extracted within a month is taken as an example and annotated according to the corresponding season (spring, summer, autumn, and winter).

[0057] Classification test: The low-dimensional feature vector of the graph data is obtained through the method model, and then the feature vector is input into the support vector machine (SVM) for classification. The classification performance is measured using two indicators, Macro-F1 and Micro-F1, which can comprehensively evaluate the performance of the model in each category in a multi-classification scenario.

[0058] Clustering test: Similarly, the low-dimensional feature vector of the graph data is obtained through the method model, and then the feature vector is clustered using the K-means algorithm. The performance of the feature representation learned by different methods on the clustering task is evaluated by using the normalized mutual information NMI and the adjusted Rand index ARI.

[0059] The larger the index, the better. The comparison results are shown in the table:

[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] From the results we can see that:

[0062] 1) In the classification task, the proposed method significantly outperforms the comparison method GAT in both Macro-F1 and Micro-F1. Specifically, the Macro-F1 of the proposed method reaches 60.78%, which is 28.8 percentage points higher than the 31.98% of GAT; on Micro-F1, the proposed method reaches 61.76%, which is 13.53 percentage points higher than the 48.23% of GAT. This shows that the proposed method can better capture the key features of multi-view graph data and generate more discriminative graph representations, thereby greatly improving the classification performance.

[0063] 2) In the clustering task, the NMI and ARI indicators of the proposed method are also significantly higher than those of GAT. Among them, the NMI of the proposed method reaches 59.86%, which is 12.45 percentage points higher than the 47.41% of GAT; in terms of the ARI indicator, the proposed method reaches 37.37%, which is 8.67 percentage points higher than the 28.70% of GAT. This further verifies the advantages of the proposed method in learning the intrinsic structure and similarity of multi-view graph data, and proves its effectiveness and adaptability in unsupervised learning scenarios.

[0064] 3) In summary, through experiments on the real-world BikeDC dataset, the method of the present invention has achieved significant performance improvements in both classification and clustering tasks, proving its superiority in extracting and aggregating multi-view graph features. Specifically, the present invention effectively improves the information over-smoothing problem existing in the existing multi-view GNN model by introducing a novel node filtering strategy and a cross-attention mechanism, and achieves more accurate and in-depth information fusion. In addition, the neighbor selection module based on unsupervised learning also enhances the applicability and flexibility of the model in an unlabeled data environment. These experimental results fully demonstrate that the present invention has achieved the intended technical purpose and provides a solution with excellent performance and strong adaptability for processing complex multi-view graph data.

[0065] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A multi-view graph neural network aggregation enhancement method based on neighbor selection, characterized in that: The following steps are involved: Step 1: Evaluate neighbor importance: Based on the adjacency matrix and feature matrix of the graph, calculate the similarity score matrix between the nodes in the graph to measure the importance of neighbor nodes; Step 2, constructing a node filtering model: introducing the concept of retention ratio, that is, the ratio of the number 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; Step 3: Balance between neighbor importance and connection information: Based on the functional relationship model, a balance factor β is introduced as a hyperparameter to weigh the importance of neighbors and the number of neighbor connections. Step 4: Information aggregation: Design an enhanced cross-view aggregation framework to achieve the mutual promotion process of intra-graph aggregation and cross-attention aggregation between graphs.

2. The multi-view graph neural network aggregation enhancement method based on neighbor selection according to claim 1 is characterized in that: The step 1 specifically comprises the following steps: Step 1.1, use Jaccard similarity on 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 between the features of two nodes; Step 1.3: Add the two similarities to obtain a comprehensive similarity score matrix that takes into account connection similarity and feature similarity.

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

4. The multi-view graph neural network aggregation enhancement method based on neighbor selection according to claim 3 is characterized in that: The step 3 specifically comprises 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 emphasis is placed on retaining neighbor nodes, and the closer the β value is to 0, the more emphasis is placed on retaining sufficient connection information; Step 3.2: Substitute the functional relationship model into the constraints of the balance factor to solve the optimal retention ratio T of each view under the corresponding β value; Step 3.3: According to the optimal retention ratio T, select the first T node pairs in the similarity score matrix, modify the original adjacency matrix, set the adjacency matrix elements of the selected node pairs to 1, and set the remaining unselected connection elements to 0, and output the filtered adjacency matrix.

5. The multi-view graph neural network aggregation enhancement method based on neighbor selection according to claim 4 is characterized in that: The step 4 specifically comprises the following steps: Step 4.1, using the enhanced adjacency matrix, the graph convolutional network (GCN) is used to propagate and aggregate node features within the graph, and the aggregated feature tensor is obtained; Step 4.2: For the obtained aggregated feature tensor, use the attention network to calculate the cumulative attention scores of other views to the current view, so as to obtain the feature matrix after inter-image aggregation; Step 4.3: Perform step 4.1 again on the feature matrix after inter-graph aggregation, that is, perform a second intra-graph aggregation, and output the feature matrix after the second intra-graph aggregation; Step 4.4: Perform the final inter-graph aggregation on the feature matrix after the secondary aggregation, convert the feature matrix output in step 4.3 into a vector form, and obtain the final feature matrix representation of the multi-view graph through the pooling operation to realize the inter-graph information fusion of the multi-view graph neural network.