A classification method based on hypergraph transformation network

Abstract

The present invention provides a classification method based on a hypergraph transformation network, which is used to solve the problem that the existing technology cannot deeply explore the high-order semantic information in heterogeneous networks, and thus cannot accurately classify. This method proposes an end-to-end hypergraph transformation network (Hypergraph Transformer Neural Networks, HGTN), which uses the communication ability between hyperedge-amplified nodes to learn high-order relationships and mine semantic information between different types of nodes. Specifically, the attention mechanism is used to assign weights to different types of hypergraphs, cascade learning of the high-order semantic information implicit in the original heterogeneous hypergraph, generate useful meta-paths, learn node embedding features in an end-to-end manner, and complete the node classification task. This method has good accuracy and universality, and is suitable for node classification tasks in heterogeneous networks such as citation networks, media networks, and social networks.

Description

Technical Field

[0001] The present invention mainly relates to the fields of graph mining, deep neural network algorithms and node partitioning, and specifically is a classification method based on a hypergraph transformation network. Background Art

[0002] In recent years, the application of deep networks to non-Euclidean data processing tasks has attracted widespread attention. The emergence of graph neural networks has made significant breakthroughs in these tasks. Figure 1 An edge can connect two vertices, indicating the relationship and message propagation between pairs of nodes. However, for complex many-to-many relationships, simple graphs will lose a lot of useful high-order information. For example, taking papers as nodes, the condition for connecting nodes is that they are written by the same author. For a simple graph, if two papers are written by the same person, a connection relationship is established. However, the reality is not so simple: the nodes in the graph may be related to multiple other nodes. Just as an author may write three or more articles, this can be considered to capture high-order dependencies between instances. Therefore, for the data structure reflected by multi-dimensional relationships, ordinary graphs have poor ability to capture the high-order information implicit in nodes or edges.

[0003] Furthermore, most network data in reality are heterogeneous, such as social networks, citation networks, and knowledge graphs. The nodes or edges of heterogeneous networks belong to different types, and the relationship between nodes implies a lot of semantic information. In traditional graph convolutional networks, this complex semantic relationship encoding is easily overlooked. In previous studies, some scholars proposed using meta-paths to construct heterogeneous graph representation models and learn the latent space representation of nodes. However, the construction of meta-paths in such methods relies too much on expert-defined weights, making it difficult to select the optimal meta-path. Other scholars automatically update meta-paths based on random walk methods, such as metapath2vec. This method extends the random walk and skip2gram frameworks to heterogeneous graphs to express the semantic and structural associations between nodes of different types. However, it still needs to provide a pre-designed meta-path solution before training.

[0004] Therefore, the learning and representation of heterogeneous networks have the following two challenges:

[0005] (1): Represents the high-order information implied by multivariate relationships. An edge in a normal graph can only connect two nodes, so some high-order information between nodes is ignored. It is necessary to enhance the communication between nodes and improve the ability to represent high-order information.

[0006] (2): Preserve rich semantic relationships between different types. In heterogeneous networks, the types of nodes are diverse, and the connection relationships contain a lot of semantic information. It is necessary to combine feature vectors with semantic information to iteratively improve the high-dimensional representation of node embedding. Summary of the invention

[0007] In order to solve the problem that the existing technology cannot deeply explore the high-order node information of heterogeneous networks and cannot accurately classify, the present invention provides a classification method of a hypergraph transformation neural network. The method proposes a hypergraph transformation neural network that can autonomously learn semantic information and represent high-order relationships, constructs heterogeneous hypergraphs for different types of relationship networks, uses the attention mechanism to generate meta-paths to represent richer semantic relationships in the hypergraph, and learns high-order node embedding of graph relationships with semantic information through a hypergraph neural network. The specific steps of the present invention are as follows:

[0008] 1. As a generalized graph, a hypergraph is a visual and intuitive mathematical expression of multi-relational data. Unlike ordinary graph structures, a hyperedge of a hypergraph can connect multiple nodes. Therefore, it can more accurately describe the relationship between objects with multi-relationships and more easily represent higher-order information of the data. The steps to obtain the hypergraph adjacency matrix A are as follows:

[0009] 1.1) Calculate the hypergraph association matrix H

[0010] Incidence Matrix Available Matrix Indicates that the vertical axis of H represents node v, and the horizontal axis represents hyperedge e, where N is the maximum number of nodes and M is the maximum number of hyperedges. The element H(v,e) in the matrix is ​​defined as:

[0011]

[0012] 1.2) Calculate the node degree matrix D v and the edge degree matrix D e

[0013] and are diagonal matrices, representing the degree matrix of the hypergraph node and the degree matrix of the hyperedge, respectively. The element D in the degree matrix of the hypergraph node i is v (i,i) is defined as:

[0014]

[0015] The degree matrix D of hyperedge i e The inner element of (i,i) is defined as:

[0016]

[0017] 1.3) Calculate the hypergraph adjacency matrix A

[0018] Adjacency Matrix of Hypergraph Represents the connection relationship of each node in the hypergraph, and A is defined as:

[0019]

[0020] 2. Obtaining the heterogeneous hypergraph adjacency matrix

[0021] Taking the paper network as an example, nodes are connected by different types of semantics (whether they belong to the same author, the same conference, or the same keyword), so τ = 3 hypergraph adjacency matrices can be constructed ∏(·) is the superposition operation.

[0022] 3. Computing Meta-Hypergraph

[0023] The meta-path P is defined as: This is actually in the node type V 1 , V l A composition relationship is defined between Represents the combination operation between relations. Taking the citation network as an example, the meta-path "APA" represents the co-authorship relationship between two authors (A) on a paper (P). The adjacency matrix is ​​represented as: A APA =A AP ×A PA , “APCPA” represents two authors (A) publishing papers (P) in the same conference (C), and the adjacency matrix is ​​represented as: A APCPA =A APC ×A CPA Therefore, the semantic information between different types of nodes can be represented by meta-paths of different lengths.

[0024] The heterogeneous hypergraph adjacency matrix Input to the cascaded attention transformation module, through the attention mechanism of this module, learn l+1 weight matrices Q of different types of graph structures, and multiply l+1 matrices Q to generate a meta-hypergraph of length l+1 To learn semantic relations,

[0025] 1.1) Attention layer: introducing convolution kernel Learning adjacency matrices for heterogeneous hypergraphs The attention parameter α can be used to obtain a single hypergraph at each layer. The weight α t , all After addition, the weight matrix Q is obtained.

[0026]

[0027] l+1 weight matrices Q can be generated through the attention layer, and the convolution kernel parameters for generating Q are different each time.

[0028] 1.2) Meta-path layer: To represent the iterative process, the adjacency matrix of the initial heterogeneous hypergraph is transformed into Recorded as 3.1.1) Perform two operations to obtain two weighted hypergraphs Q 1 (1) ,Q 2 (1) , and do matrix multiplication to get a path matrix of length 2 Later in Extract a weighted hypergraph Q from 1 (2) and Multiply to get a path matrix of length 3 And so on, we finally get the element path matrix with a length of l+1 The specific formula is:

[0029]

[0030] Here, l is generally set to the maximum value of the type, that is, l = τ.

[0031] 4. Learn node embedding feature X E :

[0032] Although the Laplacian matrix of a hypergraph can represent the relationship between nodes, it is different from that of an ordinary graph. Figure 3 As shown in Figure 1, all nodes in the same hyperedge in the hypergraph can communicate freely, and the nodes in the same hyperedge are given the same contribution: If a node has two or more hyperedges at the same time, the communication capacity between them is the weighted sum of the contributions of multiple hyperedges, recorded as Therefore, the hypergraph neural network can learn more high-order relationships based on the hyperedge contribution during propagation. Combined with random walk, the matrix θ can be established to represent the probability matrix of random walk on the hypergraph. The value of each element in the matrix is:

[0033]

[0034] in For the probability of each node being selected, the matrix normalized form of Θ can be expressed as:

[0035]

[0036] The resulting meta-path matrix It is input into the hypergraph feature embedding module at the same time as the node feature matrix X, and the node embedding features are learned through the hypergraph neural network to represent high-order relationships. The hyperedge convolution transfer formula is:

[0037]

[0038] in is the parameter to be learned in the l-th layer training.

[0039] The specific prediction process is:

[0040] enter and X to the HGNN_conv_1 layer, and learn the embedding feature X1 through the activation function Relu, input it to Dropout to solve the overfitting problem and get X2, then take the same operation: input and X2 to the HGNN_conv_2 layer, and learn the embedding feature X3 through the activation function Relu, and input it to Dropout to solve the overfitting problem to obtain the node embedding feature X E .

[0041] 5. Get the category to which the node belongs:

[0042] Next, the embedded feature X learned by the hypergraph convolution E Represents the input to the MLP layer. The MLP layer consists of multiple neurons and can be expressed by the formula f(X,W,b):

[0043]

[0044] Among them, σ(·) is a nonlinear activation function, s l is the number of neurons in the lth layer, is the connection weight between the jth neuron in layer (l) and the ith neuron in layer (l); The bias transfer formula for the i-th neuron in the (l)th layer is as follows:

[0045] X (l+1) =σ(W (l+1) X (l) +b (l+1) ).

[0046] The number of neurons in the last layer of MLP is set to the total number of categories, and the probability of each student's different score categories is output through the activation function. After training with the cross entropy loss function, the student's score is finally predicted.

[0047] Beneficial Effects

[0048] The present invention proposes a classification method based on a hypergraph transformation neural network, which is called HGTN. The method introduces the attention mechanism in Transformer to capture the semantic relationship of nodes in hyperedges, and learns more effective node representations through a hypergraph convolutional network. Compared with traditional models, the biggest advantage of HGTN is that it can capture the semantic information of high-order relationships and give nodes more effective communication capabilities. The present invention evaluates the performance of the HGTN method by performing node classification tasks on reference networks, social networks, and media networks. Experimental results show that the method proposed in the invention outperforms existing methods in performance. HGTN incorporates semantic knowledge into high-order networks for representation learning, and has considerable application potential in many tasks such as social media and personalized recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Hypergraph Transformation Layer

[0050] Figure 2 Citation network classification results

[0051] Figure 3 Media Network Classification Results

[0052] Figure 4 Social network classification results

[0053] Figure 5 (a) The accuracy of ablation experiments;

[0054] Figure 5 (b) Recall rate of ablation experiment;

[0055] Figure 5 (c) Accuracy of ablation experiments;

[0056] Figure 5 (d) Accuracy F1 coefficient of ablation experiment;

[0057] Figure 6 (a) Semantic verification on the DBLP dataset;

[0058] Figure 6 (b) Semantic verification on the ACM dataset;

[0059] Figure 6 (c) Semantic verification on the REUT dataset;

[0060] Figure 6 (d) Semantic verification on the IMDB dataset;

[0061] Figure 7 Hyperparameter Verification Results

[0062] Figure 8 (a) Accuracy of robustness experiments;

[0063] Figure 8 (b) F1 coefficient of ablation experiment;

[0064] Fig. 9 HGTN model framework DETAILED DESCRIPTION

[0065] The experimental environment of this method is: Ubuntu18.04.5, GeForce RTX 3090 24GB, CUDA=11.2, python=3.8.3, pytorch=1.7.1. The training set and test set are divided into training set and test set with a division ratio of 0.8. The main parameters are set as: Epoch=500, learning rate=0.0005, weight decay=0.001.

[0066] Obtain the paper data and make the data set into two parts: node feature X and graph structure. The node feature is the ONE-HOT code of the paper title, and the graph structure describes the topological structure of the node. The association matrix can be used as a matrix Indicates that the vertical axis of H represents node v, and the horizontal axis represents hyperedge e, where N is the maximum number of nodes and M is the maximum number of hyperedges. The element H(v,e) in the matrix is ​​defined as:

[0067]

[0068] calculate and The element D in the degree matrix of hypergraph node i v (i,i) is defined as:

[0069]

[0070] The degree matrix D of hyperedge i e The inner element of (i,i) is defined as:

[0071]

[0072] Adjacency Matrix of Hypergraph Represents the connection relationship of each node in the hypergraph, and A is defined as:

[0073]

[0074] Nodes are connected by different types of semantics (whether they belong to the same author, the same conference, or the same keyword), so τ = 3 hypergraph adjacency matrices can be constructed ∏(·) is the superposition operation.

[0075] Introducing convolution kernel Learning adjacency matrices for heterogeneous hypergraphs The attention parameter α can be used to obtain a single hypergraph at each layer. The weight α t , all The added weight matrix Q is obtained. l+1 weight matrices Q can be generated through the attention layer. The convolution kernel parameters of each generation of Q are different. The l+1 weight matrices Q are multiplied to obtain the meta-path matrix Here, l is generally set to the maximum value of the type, that is, l = τ.

[0076] The resulting meta-path matrix It is input into the hypergraph feature embedding module at the same time as the node feature matrix X. The node embedding features are learned through the hypergraph neural network to represent high-order relationships. The hypergraph feature embedding module includes the first hyperedge convolution HGNN_conv_1, the second hyperedge convolution HGNN_conv_2, Dropout, activation function Relu, and fully connected layer Linear. The specific connection relationship is HGNN_conv_1→Relu→Dropout→HGNN_conv_2→Relu→Dropout→Linear.

[0077] enter and X to the HGNN_conv_1 layer, and learn the embedding feature X1 through the activation function Relu, input it to Dropout to solve the overfitting problem and get X2, then take the same operation: input and X2 to the HGNN_conv_2 layer, and learn the embedding feature X3 through the activation function Relu, and input it to Dropout to solve the overfitting problem to obtain the node embedding feature X E The obtained embedding feature X E Input to the MLP layer, which consists of linear neurons and activation functions, and performs SGD gradient descent iterative training through the cross entropy loss function to finally get the category to which the node belongs

[0078] The present invention conducts experiments on six datasets in the fields of not only citation networks, but also media networks and social networks. However, all datasets can be represented as two inputs: node features and topological relationships. At the same time, several advanced methods are selected from three types of models: graph relationship representation, graph neural network, and hypergraph neural network. The performance of these methods on heterogeneous network analysis tasks is compared from multiple perspectives.

[0079] The experiment selected six datasets in different fields:

[0080] DBLP: The paper citation network extracted by the DBLP website contains three node relationship graphs: co-authorship (two authors jointly publish the same paper), joint conference attendance (two authors jointly attend the same conference), and common keywords (two authors use the same keyword).

[0081] ACM: The citation network of papers in the ACM database. It contains three node relationship graphs: common authors (two papers contain the same author), common conference participants (two papers participated in the same conference), and common keywords (the same keyword appears in two papers).

[0082] CITE: Paper citation network, describing the citations between papers in the world's top conferences. Construct 2 knn node relationship graphs based on the features.

[0083] IBDM: The data comes from the IMDB movie introduction and rating website. It contains three node relationship graphs: co-actors (two movies contain the same actor), co-directors (two movies contain the same director), and co-years (two movies were released in the same year).

[0084] REUT: Contains many short news and their corresponding topics, published by Reuters in 1986. It is a simple and widely used text classification dataset. 2 knn node relationship graphs are constructed based on the features.

[0085] STUD: The data is collected from student activity information of a university, describing the student social network composed of multiple behavioral associations. According to different behaviors, it can be divided into 7 node relationship graphs.

[0086] The present invention selects 6 advanced methods in the three fields of graph relationship representation, graph neural network, and hypergraph neural network, and compares them with the present invention. The 6 methods are introduced as follows:

[0087] DeepWalk: A graph data mining algorithm that combines the random walk and word2vec algorithms. The algorithm can learn the hidden information of the network and represent the nodes in the graph as a vector containing potential information.

[0088] LINE: A graph embedding method that uses a breadth-first random walk to obtain contextual information and learns network representations through first-order and second-order similarities.

[0089] GCN: A graph neural network using convolution operations, which can be applied to graph embedding feature representation learning;

[0090] GAT: Graph neural network, which aggregates neighbor nodes through self-attention mechanism and realizes adaptive matching of weights of different neighbors;

[0091] HGNN: A hypergraph neural network using convolutional operations that can represent and learn high-order embedding features;

[0092] GTN: Graph Transformation Network, which updates the graph by generating meta-paths and uses convolution to represent the node features of the graph structure.

[0093] In order to more comprehensively evaluate the performance of HGTN, this paper selects four indicators: accuracy, F1 score, precision, and recall for comparative experiments. The experimental graphics card is NVIDIA GeForce RTX 3090, 24GB G6X video memory.

[0094] The heterogeneous hypergraph composed of the reference network is input into HGTN for node classification. Figure 2 As shown in the figure, HGTN outperforms the baseline method in most cases. It should be noted that in the ACM dataset, HGTN and GTN have similar results, both as high as 0.99. This is because both methods take into account semantic information, making the classification effect close to 100%. However, it was found that the F1 coefficient of HGTN in the CITE dataset is not high. Analysis shows that the graph in CITE is constructed using KNN, which is more affected by noise and has less semantic information that can be mined. The results of the media and social network node classification tasks are shown in Figure 2. Figure 3 Figure 4 As shown in the figure, HGTN outperforms HGNN and GTN in all evaluation indicators. The average improvement is about 3%, which proves that HGTN captures the high-order semantic information of nodes and has strong universality.

[0095] In order to verify the solution of the two challenges described in the introduction, two sets of ablation experiments are set up: HGTN (H) and HGTN (T). HGTN (H) only targets the first challenge, that is, it only considers high-order relations and does not consider semantic relations. HGTN (T) only considers semantic relations and does not consider high-order relations. The present invention analyzes the two ablation methods and HGTN in terms of evaluation indicators, and the results are as follows Figure 5 As shown. Experiments show that the present invention improves both challenges in data classification tasks. The maximum improvement is 10% and the average improvement is 4%.

[0096] In order to analyze the improvement of semantic relationship mining on experimental results, the present invention gradually introduces different types of hypergraphs to generate meta-paths on four datasets: DBLP, ACM, REUT and IMDB. Figure 6 As shown in the figure, it can be seen that with the introduction of different types of hypergraphs containing semantic relationships, the evaluation index of the model also improves. This shows that the newly generated meta-path has indeed mined useful semantic information and improved the accuracy.

[0097] The addition of channels can help better capture semantic information, but it will also increase the loss of computing time. Since the GTN model also considers the semantic impact, this section further analyzes the corresponding ACC, F1 and computing time of GTN and HGTN with different channel numbers c (1, 2, 3, 4, 5) Figure 7 ) to determine the optimal parameters. To ensure fairness, the training parameters are kept consistent, the training epoch is 400, the learning rate is 0.001, and the average is calculated over multiple experiments. The experimental results are shown in Figure 7 As shown in the figure, it is proved that under the same number of channels, HGTN's acc, f1 and time indicators are better than GTN. When c=4, the model has the highest accuracy and the fastest running speed is 34s.

[0098] In order to verify the robustness of HGTN, the present invention uses the noise attack model implied by behavioral features in the STUD dataset. This dataset determines a variety of student behavioral features based on the school's work and rest time and the function of each location, and constructs a hypergraph based on feature similarity. However, this dataset is subject to noise interference. The main limitation is that measuring social relationships using feature similarity between nodes may not be accurate due to noise and outliers. Another problem is that it is impossible to determine how many neighbors the central node should be connected to, because the size of the nearest neighbors may affect the performance of hypergraph learning. Therefore, the present invention assumes that changes in the number of neighbors of the central node k are interference. In the experiment, the present invention analyzed the results of different k values ​​in the evaluation index, and the results are as follows. Figure 8 As shown in Figure 2, under noise interference, the variation range of ACC and F1 is small, and the difference between the best and the worst is about 3%, which proves that HGTN has good robustness.

Claims

1. A classification method based on a hypergraph transformation neural network, which integrates high-order associations between nodes and semantic information of different types of edges to perform node classification, characterized by: The paper is used as a node, and the one-hot encoding of the paper title is used as the node feature. The nodes are connected by different types of semantics. Different types of semantics include whether they belong to the same author, the same conference, or the same keyword. One type of semantics connects multiple nodes to form a hypergraph, which is represented by the adjacency matrix A. Multiple hypergraphs are formed according to different types of semantics, and the adjacency matrix A is used to represent them. t It is represented by, where t = 1, 2, ..., τ, τ is the total number of semantic types, and these τ hypergraphs are superimposed together to represent the heterogeneous hypergraph adjacency matrix Cascaded attention transformation module input heterogeneous hypergraph Generate a meta-hypergraph through the attention layer and meta-path layer The hypergraph feature embedding module inputs the meta-hypergraph and node features X, output node feature embedding X E ; Node category prediction module input node embedding X E Get the node prediction label after the MLP layer The hypergraph adjacency matrix A is specifically obtained as follows: 1.1) Calculate the hypergraph association matrix H: Incidence Matrix Available Matrix Indicates that the vertical axis of H represents the node v, and the horizontal axis represents the hyperedge e, where N is the maximum number of nodes, M is the maximum number of hyperedges, and the element H(v,e) in the matrix is ​​defined as: 1.2) Calculate the node degree matrix D v and the edge degree matrix D e : and Both are diagonal matrices, representing the degree matrix of the hypergraph node and the degree matrix of the hyperedge, respectively. The element D in the degree matrix of the hypergraph node i is v (i,i) is defined as: The degree matrix D of hyperedge i e The inner element of (i,i) is defined as: 1.3) Calculate the hypergraph adjacency matrix A Adjacency Matrix of Hypergraph Represents the connection relationship of each node in the hypergraph, and A is defined as: The heterogeneous hypergraph adjacency matrix The specific acquisition process is as follows: For the paper network, nodes are connected by τ different types of semantics, so τ hypergraph adjacency matrices are constructed ∏(·) is a superposition operation.

2. The classification method based on hypergraph transformation neural network according to claim 1, characterized in that: The specific operation process of the cascaded attention transformation module is as follows: The heterogeneous hypergraph adjacency matrix Input to the cascaded attention transformation module, through the attention mechanism of this module, learn l+1 weight matrices Q of different types of graph structures, and multiply l+1 matrices Q to generate a meta-hypergraph of length l+1 To learn semantic relations, 1.1) Generate l+1 weight matrices Q through the attention layer, where the generation process of any weight matrix Q is: Introducing convolution kernel Learning adjacency matrices for heterogeneous hypergraphs The attention parameter α can be used to obtain a single hypergraph at each layer. The weight α t , all After addition, we get the weight matrix Q. The convolution kernel parameters for generating Q are different each time; 1.2) Use the meta-path layer to obtain a meta-path matrix of length l+1 The specific process is as follows: In order to represent the iterative process, the adjacency matrix of the initial heterogeneous hypergraph is Recorded as According to the formula (5) in 1.1) of claim 4, Perform two operations to obtain two weighted hypergraphs Q1 (1) ,Q2 (1) , and do matrix multiplication to get a path matrix of length 2 Two of the operations use different convolution kernel parameters; then Formula (5) is used to extract a weighted hypergraph Q1 (2) and Multiply to get a path matrix of length 3 And so on, we finally get the element path matrix with a length of l+1 The specific formula is: Among them, α l,t Represents the lth convolution kernel w φ Obtained The weight of the t-th layer single hypergraph, l is generally set to the maximum value of the type, that is, l = τ.

3. The classification method based on hypergraph transformation neural network according to claim 1, characterized in that: The specific operation process of the hypergraph feature embedding module is as follows: The resulting meta-path matrix It is input into the hypergraph feature embedding module at the same time as the node feature matrix X. The node embedding features are learned through the hypergraph neural network to represent high-order relationships. The hypergraph feature embedding module includes the first hyperedge convolution HGNN_conv_1, the second hyperedge convolution HGNN_conv_2, Dropout, activation function Relu, and fully connected layer Linear. The specific connection relationship is HGNN_conv_1→Relu→Dropout→HGNN_conv_2Relu→Dropout→Linear, where the hyperedge convolution transfer formula is: in are the parameters to be learned during training, The specific prediction process is: enter and X to the HGNN_conv_1 layer, and learn the embedding feature X1 through the activation function Relu, input it to Dropout to solve the overfitting problem and get X2, then take the same operation: input and X2 to the HGNN_conv_2 layer, and learn the embedding feature X3 through the activation function Relu, and input it to Dropout to solve the overfitting problem to obtain the node embedding feature X E .

4. The classification method based on hypergraph transformation neural network according to claim 1, characterized in that: The node category prediction module has the following specific operation process: The obtained embedding feature X E The input is sent to the MLP layer, which is composed of linear neurons and activation functions. After the cross entropy loss function, the SGD gradient descent iterative training is performed to finally obtain the category to which the paper belongs.

5. The classification method based on hypergraph transformation neural network according to claim 1, characterized in that: The nodes are further movies, news, or student activity information.

Images