A node classification method based on fusion of high communication signals and feature correlation

Abstract

The application discloses a node classification method based on fusion of high communication signals and feature correlation, which comprises the following steps: S1, processing data and constructing network data; S2, constructing a feature correlation graph by using a relaxed minimum spanning tree; S3, designing corresponding low-pass filters and high-pass filters; S4, applying the low-pass filters and the high-pass filters to an original topology graph and the feature correlation graph to obtain high communication signals and low communication signals of node features in the two graphs, and designing two different fusion modes to fuse the two different signals of the two graphs to obtain representations of different nodes in the two graphs; S5, obtaining a final node representation; and S6, dividing nodes with the same label into a class. The application introduces the high communication signals to make the feature difference of different label nodes larger, and retains the low communication signals to make the feature difference of the same label nodes smaller, thereby improving the node classification accuracy from two aspects.

Description

Technical Field

[0001] This invention belongs to the technical field of graph learning, and particularly relates to a node classification method based on fusing high-pass signals and feature correlations. Background Technology

[0002] Graphs, representing entities and their relationships, are ubiquitous in the real world, appearing in social networks, point clouds, transportation networks, knowledge graphs, and molecular structures. In social networks, each entity node represents an actual social user, and edges represent connections between these users. In citation networks, each entity node represents a specific academic article, and edges represent citation relationships between two articles. Therefore, nodes represent different meanings in network data with different structures, and studying node classification has significant applications in many real-world scenarios.

[0003] Recently, many node classification methods have utilized popular deep learning techniques to learn a better node representation. However, graph structures are non-Euclidean, and graph neural networks (GNNs) have emerged to address this challenge, enabling deep learning to learn a good representation in graph data and achieving great success in many applications, such as node classification, graph classification, and link prediction. However, most current GNN models still have problems. On the one hand, they only consider the first- or second-order neighbor information of the target node, ignoring the information of nodes with similar features but distant locations in the topological graph. On the other hand, most GNN models are considered low-pass filters, ignoring the high-pass signals of node features, which to some extent affects the node representation. Summary of the Invention

[0004] To address the problems in the background technology, this invention proposes a node classification method based on fusing high-pass signals and feature correlations, the specific scheme of which is as follows:

[0005] A node classification method based on fusing high-pass signals and feature correlations includes the following steps:

[0006] S1. Process data and construct network data;

[0007] S2. Calculate the feature correlation between node features based on cosine similarity, and then use relaxed minimum spanning tree to construct feature correlation graph;

[0008] S3. Based on graph signal theory and graph wavelet neural network model, design the corresponding low-pass and high-pass filters.

[0009] S4. Apply low-pass and high-pass filters to the original topology diagram and feature correlation diagram to obtain high-pass and low-pass signals of node features in the two diagrams, and design two different fusion methods to fuse the two different signals to obtain the representations of different nodes in the two diagrams.

[0010] S5. Use an attention mechanism to adaptively fuse representations from two different types of graph nodes to obtain the final node representation;

[0011] S6. Input the final node representation into the node classification model to classify the nodes, predict the classification label of the nodes with unknown labels, and classify the nodes with the same label into one class.

[0012] Specifically, step S1 includes:

[0013] S11. Treat the specific entities in the actual network as individual nodes, the connections between entities as edges of nodes, the category information of each entity as the label of the node, and the attributes of the node as the features of the node.

[0014] S12. Construct a network G = {V, E, X} from the processed entity data, where V represents the set of nodes in the network, which has n nodes, and is denoted as V = {V0, V1, V2, ... V}. n-1 The set of m edges in a network is represented by E = {e0, e1, e2, ..., e}. n-1 The statement indicates that in a network, if two entity nodes a and b are directly connected, then there exists an edge e = {a, b} ∈ E between the two nodes; X represents the n×k node features, indicating that there are n nodes in the network structure, each with k-dimensional feature information, and X is used to define these features. i This represents the information of the attributes of the i-th node.

[0015] Specifically, step S2 includes the following steps:

[0016] S21. Use cosine similarity distance to calculate the distance between node features. The similarity calculation between nodes A and B is defined as follows:

[0017]

[0018] S22. Based on the similarity matrix S, construct the feature correlation map using the relaxed minimum spanning tree algorithm. The specific steps are as follows:

[0019]

[0020] in, Let d(i,j) represent the adjacency matrix of the newly constructed feature correlation graph, and let d(i,j) represent the distance between the direct connections of node i and node j. k ),d(j,j k ) represent the distances between node i and node j and their nearest k-th neighbor, respectively, where γ is a hyperparameter controlling the information weights of the minimum spanning tree path and the nearby neighbor nodes of the two nodes, mlink ij This represents the sum of the path weights that have the highest weight among the multiple interconnected paths of the minimum spanning tree for the node pair (i,j).

[0021] Specifically, step S3 includes:

[0022] Design convolution kernels I-∈1*Λ1 for the low-pass filter and I+∈1*Λ1 for the high-pass filter separately for the original topology diagram, where ∈1 are trainable parameters of the low-pass and high-pass filters, and their values ​​are limited to [0,1]. The convolution kernel of the low-pass filter is written as g. θ (λ i )=1-∈1*λ i .

[0023] Specifically, step S4 includes:

[0024] S41. Generate corresponding low-pass and high-pass signal representations for the original topology diagram and feature correlation diagram, Z. L1 Z H1 Z represents the low-pass and high-pass signals of the original topology diagram, respectively. L2 Z H2 The low-pass and high-pass signals of the feature correlation map are represented respectively, and their specific formulas are shown below:

[0025]

[0026]

[0027]

[0028]

[0029] in, and Wavelet basis and inverse wavelet basis representing the original topological structure graph, and Representing the wavelet basis and inverse wavelet basis of the feature correlation plot;

[0030] S42. For the original topological graph, a trainable parameter α is set to fuse the node feature signals of the graph, and the formula is as follows:

[0031] Z1=α*Z L1 +(1-α)*Z H1

[0032] Z1 represents the node representation on the original topology graph;

[0033] For the feature correlation map, when fusing high-pass and low-pass signals, in addition to setting a trainable parameter r to fuse these two different signals, an extra hyperparameter β is set to suppress the high-pass signal. The specific fusion method is as follows:

[0034] Z2=r*Z L2 +(1-r)*β*Z H2

[0035] Z2 represents the node representation on the feature correlation graph.

[0036] Specifically, step S5 includes:

[0037] S51. Obtain the representations of the original topological structure graph and the feature correlation graph. Introduce an attention mechanism to fuse the representations of the two different types of graphs to obtain the final node representation. The specific details are shown below:

[0038] (k1,k2)=attention(Z1,Z2)

[0039] Where k1 and k2 represent the weights learned from the original topological structure graph and the feature correlation graph during the training process, respectively.

[0040] S52. After obtaining the weights, perform linear summation to obtain the final node representation Z, as shown in the following formula:

[0041] Z = (K1*Z1 + K2*Z2).

[0042] Specifically, step S6 is as follows:

[0043] After obtaining the final node representation, a multilayer perceptron and the nonlinear transformation softmax() function are applied to predict the labels of unknown labeled nodes, thus classifying the nodes. The formula is shown below:

[0044] Final = softmax(W*Z+B).

[0045] The beneficial effects of this invention are as follows: considering the role of feature correlation, constructing a feature correlation map can cleverly obtain information about nodes that are far apart in the original topology map but have similar features, thus enriching the representation of nodes; introducing a high-pass signal can increase the feature differences between nodes with different labels, while retaining a low-pass signal can reduce the feature differences between nodes with the same label, thereby improving the accuracy of node classification from two aspects. Attached Figure Description

[0046] Figure 1 This is a flowchart of a node classification method based on fusing high-pass signals and feature correlation proposed in this invention. Detailed Implementation

[0047] like Figure 1 As shown, a node classification method based on fusing high-pass signals and feature correlations is presented below. The invention is further illustrated using the Citeseer citation network as a specific example, and includes the following steps:

[0048] S1. Process the data citation network Citeseer and construct the network data; the citation network Citeseer contains 3312 articles, with 4660 citation or reference relationships between articles, and each article contains a title and abstract.

[0049] Specifically, step S1 includes:

[0050] S11. Treat each paper in the Citeseer citation network as a node in the network and number them sequentially from 0, 1, 2, 3, ..., 3311. Treat citation or cited relationships as edges in the network, and number the article categories 0, 1, 2, 3, 4, 5 as labels for each node. After removing stop words and low-frequency words (less than 10 occurrences) from the title and abstract of each article, 3703 words are obtained, averaging 32 words per document, which are used as the article's attribute information.

[0051] S12. Based on the processed data, construct a network G = {V, E, X}, where V represents a set of 3312 nodes, denoted as V = {v0, v1, ... v...}. 3311 Let E represent the set of 4660 edges in the network. If there is an edge connecting nodes a and b, then (a, b) ∈ E, denoted as E = {e1, e2, ..., e}. 4660 Let X be a 3312×3703 dimensional matrix, where the i-th row represents the information of the i-th node. i express.

[0052] S2. Calculate the feature correlation between node features based on cosine similarity, and then use relaxed minimum spanning tree to construct feature correlation graph;

[0053] The specific steps are as follows:

[0054] S21. Use cosine similarity distance to calculate the distance between node features, indicating the degree of similarity between node features. Formally, the similarity calculation between nodes A and B is defined as follows:

[0055]

[0056] S22. Based on the similarity matrix S, construct the feature correlation map using the relaxed minimum spanning tree algorithm. The specific steps are as follows:

[0057]

[0058] in, Let d(i,j) represent the adjacency matrix of the newly constructed feature correlation graph, and let d(i,j) represent the distance between the direct connections of node i and node j. k ),d(j,j k ) represent the distances between node i and node j and their nearest k-th neighbor, respectively, where γ is a hyperparameter controlling the information weights of the minimum spanning tree path and the nearby neighbor nodes of the two nodes, mlink ij The sum of the path weights that represents the largest weight among the multiple interconnected paths of the minimum spanning tree for the node pair (i,j).

[0059] S3. Based on graph signal theory and graph wavelet neural network model, design the corresponding low-pass and high-pass filters.

[0060] Because high-pass signals can preserve the differences between node features, while low-pass signals can preserve the similarities between node features, the difference information can make the differences between node features with different labels larger during the learning process, while the similarity information can make the differences between node features with the same label smaller during the learning process. The specific steps are as follows:

[0061] Design the convolution kernel I-∈1*Λ1 for the low-pass filter and the convolution kernel I+∈1*Λ1 for the high-pass filter based on the original topology graph, where ∈1 are trainable parameters of the low-pass and high-pass filters, with values ​​limited to [0,1]. Rewrite the convolution kernel of the low-pass filter as g θ (λ i )=1-∈1*λ i When λ i When = 0, g θ (λ i ) = 1, as λ i continuously increasing, g θ (λ i The value of λ continuously decreases, indicating that it plays a role in suppressing high-pass signals; for the convolution kernel of the high-pass filter, as λ... i continuously increasing, g θ (λ i The value of ) is constantly increasing, indicating that it amplifies the high-pass signal.

[0062] S4. Apply low-pass and high-pass filters to the original topology diagram and feature correlation diagram to obtain high-pass and low-pass signals of node features in the two diagrams, and design two different fusion methods to fuse the two different signals to obtain the representations of different nodes in the two diagrams.

[0063] The specific steps are as follows:

[0064] S41 generates corresponding low-pass and high-pass signal representations for the original topology diagram and feature correlation diagram, Z. L1 Z H1 Z represents the low-pass and high-pass signals of the original topology diagram, respectively. L2 Z H2 The low-pass and high-pass signals of the feature correlation map are represented respectively, and their specific formulas are shown below:

[0065]

[0066]

[0067]

[0068]

[0069] in, and Wavelet basis and inverse wavelet basis representing the original topological structure graph, and Representing the wavelet basis and inverse wavelet basis of the feature correlation plot;

[0070] S42. For the original topological graph, a trainable parameter α is set to fuse the node feature signals of the graph, and the formula is as follows:

[0071] Z1=α*Z L1 +(1-α)*Z H1

[0072] Z1 represents the node representation on the original topology graph.

[0073] For the feature correlation map, given that most of the neighboring nodes of the target node have similar features, and considering that introducing too much high-pass signal may affect the node representation, when fusing high-pass and low-pass signals, in addition to setting a trainable parameter r to fuse these two different signals, an extra hyperparameter β is set to suppress the high-pass signal. The specific fusion method is as follows.

[0074] Z2=r*Z L2 +(1-r)*β*Z H2

[0075] Z2 represents the node representation on the feature correlation graph.

[0076] S5. Adaptively fuse representations from two different types of graph nodes using an attention mechanism to obtain the final node representation; the specific steps are as follows:

[0077] S51. Obtain the representations of the original topological structure graph and the feature correlation graph. Introduce an attention mechanism to fuse the representations of the two different types of graphs to obtain the final node representation. The specific details are shown below:

[0078] (k1,k2)=attention(Z1,Z2)

[0079] Where k1 and k2 represent the weights learned from the original topological structure graph and the feature correlation graph during the training process, respectively.

[0080] S52. After obtaining the weights, perform linear summation to obtain the final node representation Z, as shown in the following formula:

[0081] Z = (K1*Z1 + K2*Z2)

[0082] S6. Input the final node representation into the node classification model to classify the nodes, predict the classification label of nodes with unknown labels, and group nodes with the same label into the same class. The specific steps are as follows:

[0083] After obtaining the final node representation, a multilayer perceptron and the nonlinear transformation softmax() function are applied to predict the labels of unknown labeled nodes, thus classifying the nodes. The formula is shown below:

[0084] Final = softmax(W*Z+B).

[0085] This invention introduces feature correlation and high-pass signal to improve node representation capability, fully acquires node information with similar features but far apart locations in the original topology graph, and also fully acquires the difference information between nodes with different labels, so that the node representation has higher discriminativeness. Both improve node representation to a certain extent and improve the accuracy of node classification.

[0086] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A node classification method based on fusing high-pass signals and feature correlations, characterized in that, Includes the following steps: S1. Process data and construct network data; Step S1 specifically includes: the network data is a paper citation network; S11. Treat the papers in the paper citation network as individual nodes, the paper citation relationships as edges of the nodes, the paper's subject and topic category as labels for the nodes, and the paper's text content as features of the nodes. S12. Construct a network from the processed network data. ,in Let n be the set of paper nodes in the paper network, and denote it as n nodes. For; the set of m edges of the network. This indicates that two paper nodes are assumed in the network. If there is a direct connection, then there is an edge between the two nodes. ; It means The node features represent the n nodes in the paper's network structure, with each node possessing k-dimensional feature information, and are used... This represents the paper content information of the i-th node; S2. Calculate the feature correlation between the features of the paper nodes based on the cosine similarity distance, and then use the relaxed minimum spanning tree to construct the feature correlation map of the paper nodes. S3. Based on graph signal theory and graph wavelet neural network model, design low-pass and high-pass filters for citation network data. S4. Apply low-pass and high-pass filters to the original topology graph and feature correlation graph of the citation network to obtain high-pass and low-pass signals of the paper node features in the two graphs, and design two different fusion methods to fuse the two graphs into two different signals to obtain the representations of different paper nodes in the two graphs. S5. Use an attention mechanism to adaptively fuse representations from two different types of graph nodes to obtain the final node representations in the paper. S6. Input the final node representation into the node classification model to classify the nodes, predict the classification label of nodes with unknown labels, and classify nodes with the same label into one category. Step S6 is as follows: After obtaining the final node representations, a multilayer perceptron and the nonlinear transformation softmax() function are applied to predict the labels of unknown labeled nodes, thus classifying the nodes. It is the weight matrix of the linear transformation. This represents the bias vector. The formula for representing the predicted paper node label is shown below: ； Step S4 specifically includes: S41. Generate corresponding low-pass and high-pass signal representations for the original topology graph and feature correlation graph of the citation network, where H represents the initial features of the paper nodes in the original topology graph and feature correlation graph of the citation network, and Z... L1、 Z H1 Z represents the low-pass and high-pass signals of the paper nodes in the original topology diagram of the citation network, respectively. L2、 Z H2 The low-pass and high-pass signals of the paper nodes in the feature correlation graph are represented respectively, and their specific formulas are shown below: Z L1 = 1[ ] 1 -1 H Z H1 = 1[ ] 1 -1 H Z L2 = 2[ ] 2 -1 H Z H2 = 2[ ] 2 -1 H in, 1 and 1 -1 Wavelet basis and inverse wavelet basis representing the original topological structure graph of the citation network. 2 and 2 -1 Representing the wavelet basis and inverse wavelet basis of the feature correlation plot; S42. Set trainable parameters for the original topology of the citation network. The formula for fusing the node feature signals of this graph is as follows: Z1= WITH L1 +(1- ) WITH H1 Z1 represents the node representation on the original topology graph of the citation network; For the feature correlation map, when fusing high-pass and low-pass signals, a trainable parameter is set. In addition to fusing these two different signals, an extra hyperparameter was set. The specific fusion method for suppressing high-pass signals is as follows: Z2= WITH L2 +(1- ) WITH H2 Z2 represents the node representation on the feature correlation graph.

2. The node classification method based on the fusion of high-pass signals and feature correlations according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21. Use cosine similarity distance to calculate the distance between features of nodes in the paper. The similarity calculation between nodes A and B in the paper is defined as follows: S22. Based on the similarity matrix S, the relaxed minimum spanning tree algorithm is applied to construct the feature correlation map of the paper nodes. The specific steps are as follows: in, The adjacency matrix represents the feature relevance graph of the newly constructed paper nodes. and These represent two paper nodes respectively. This represents the distance between paper node i and paper node j that are directly connected. and represent the distances between paper node i, paper node j and its nearest k-th neighbor paper node, respectively. It is a hyperparameter that controls the information weights of the minimum spanning tree path and the neighboring paper nodes of its two paper nodes. This represents the sum of the path weights with the highest weight among the multiple interconnected paths of the minimum spanning tree for the paper node pair (i,j).

3. The node classification method based on the fusion of high-pass signals and feature correlations according to claim 1, characterized in that, Step S3 specifically includes: Design a separate convolution kernel for the low-pass filter in the original topology graph of the citation network. Convolution kernel of high-pass filter ,in It concerns the trainable parameters in low-pass and high-pass filters, whose values ​​are limited to... The convolution kernel of the low-pass filter is written as: .

4. The node classification method based on the fusion of high-pass signals and feature correlations according to claim 1, characterized in that, Step S5 specifically includes: S51. Obtain the topological structure diagram and feature correlation diagram of the original citation network. An attention mechanism is introduced to fuse the representations of two different types of graphs to obtain the final representation of the nodes in the paper. The specific formula is as follows: in These represent the weights learned during the training process from the original topology graph and the feature correlation graph of the citation network, respectively. S52. After obtaining the weights, perform linear summation to obtain the final paper node representation. The formula is as follows: 。

Images