Graph clustering method based on representation fusion
By introducing a representation fusion mechanism in the graph clustering method, combining low-pass filters, improved automatic encoders and graph convolutional networks, the oversmooth problem in the prior art is solved, and the generated representations are more distinctive and clustering performance is improved.
Patent Information
- Application Number
- CN202510073915.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-17
- Publication Date
- 2025-06-20
AI Technical Summary
The existing clustering methods based on graph convolution networks (GCN) and autoencoders (AE) often encounter the problem of smoothing in the sample encoding stage, and lack representation fusion mechanisms, resulting in the lack of distinction between the generated node representations, limiting the performance of the clustering algorithm.
A graph clustering method based on representation fusion is proposed. The low-pass filter module denoised the graph data. The encoding module uses an improved automatic encoder (IAE) and graph convolution network (GCN) to encode the graph data. The representation fusion module uses a dynamic representation fusion mechanism to fuse different representations. The dual supervision module uses a high confidence distribution, self-supervision and mutual supervision to jointly supervise the clustering and encoding process.
The representation generated by the representation fusion mechanism is more comprehensive and accurate, avoiding excessive smoothing problems, improving clustering performance, making the generated node representation more distinctive and more suitable for downstream clustering tasks.
Smart Images
Figure CN120180158A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of graph clustering and relates to a graph clustering method based on representation fusion. Background Art
[0002] As an emerging clustering technology, graph clustering has received extensive attention in recent years. Its core goal is to reveal the potential graph structure of data and effectively divide nodes into different groups. However, existing clustering methods based on graph convolutional networks (GCNs) and autoencoders (AEs) often encounter the problem of over-smoothing during the sample encoding stage and lack a representation fusion mechanism to fuse the representations generated by GCNs and AEs, resulting in the generated node representations lacking distinctiveness and limiting the performance of the clustering algorithm.
[0003] To address these problems, the present invention provides a graph clustering method based on representation fusion. This method first denoises the graph data and then fuses different representations through a representation fusion mechanism. The generated representations fully reflect the attributes and structure of the graph and are more suitable for downstream clustering tasks. Summary of the Invention
[0004] In view of this, the purpose of the present invention is to provide a graph clustering method based on representation fusion to solve the problem of over-smoothing that often occurs in existing clustering algorithms during clustering. Through a representation fusion mechanism, the attribute information and structure information of graph data are fully fused. The generated representations retain more features and structures, are more discriminative, avoid the over-smoothing problem, and are also more suitable for downstream clustering tasks.
[0005] The technical solution adopted by the present invention to achieve the above-mentioned invention purpose is as follows:
[0006] This method includes a low-pass filter module, an encoding module, a representation fusion module, and a dual supervision module.
[0007] The low-pass filter module constructs a parameter-free low-pass filter based on a graph convolutional network and filters high-frequency noise in the attribute information using the attribute information and structure information of the graph. The filtered attribute information is more suitable for clustering.
[0008] The encoding module encodes the attribute information and structure information of graph data using an improved autoencoder (IAE) and a graph convolutional network (GCN) to generate corresponding embedded representations H and Z.
[0009] The representation fusion module uses a dynamic representation fusion mechanism to finely process the attribute and structure information extracted from the improved autoencoder (IAE) and the graph convolutional network (GCN) to obtain a more comprehensive and accurate representation structure.
[0010] The dual supervision module optimizes the model training and clustering results by extracting high-confidence distributions and jointly supervising the clustering and encoding processes through self-supervision and mutual supervision.
[0011] It is characterized by the following steps:
[0012] S1: Data collection and processing: Select a publicly available dataset as the initial data, and perform low-pass filtering on the attribute information of the graph data through a low-pass filter module to filter out high-frequency noise and generate filtered attribute information;
[0013] S2: Representation learning: Encode the attribute information and structural information through IAE and GCN to generate corresponding embedding representations;
[0014] S3: Representation fusion: Dynamically fuse the representations from IAE and GCN through a representation fusion module to learn a more suitable representation for clustering, avoid the over-smoothing problem, and improve the clustering performance.
[0015] S4: Iterative training: Combine the generated representations above to optimize the model training, and supervise the representation learning and clustering processes through the dual supervision module to improve the clustering results.
[0016] S5: Effect evaluation: After training the model according to the training set, test the actual effect of the model with the test set, and evaluate the clustering ability of the model according to common model performance evaluation metrics.
[0017] Furthermore, in step S1:
[0018] S1.1: Download common graph clustering datasets: Cite, Cora, ACM, Pubmed from the Cornell University Digital Library.
[0019] S1.2: The downloaded dataset file format is npy file, and write a script program to change it to txt file for model training.
[0020] S1.3: The description of the dataset is as follows: Given an attributed graph G=(V, E, X), where is the vertex set, n is the number of nodes in the graph, E is the set of edges in the graph, and X =
[0021] [x1, x2, …, x n T is the attribute information of the graph. The topological structure of the graph can be represented by the adjacency matrix If v ij = 1, it means there is an edge between node v i and node v j , otherwise it means there is no edge between the corresponding nodes. represents the degree matrix of A, where Denote the degree of node v i .
[0022] S1.4: Construct a parameterless low-pass filter, and the form of the filter is shown in Equation 1
[0023]
[0024] where F is the filter, I is the identity matrix, is the symmetric normalized Laplacian matrix, is the adjacency matrix after normalization, is the degree matrix after normalization
[0025] S1.5: Perform low-pass filtering on the attribute information of the graph data, and the formula is shown in Equation 2
[0026]
[0027] where t represents the stacking layer number of the filter, represents the attribute information after filtering out high-frequency noise
[0028] Furthermore, in step S2:
[0029] S2.1: Take the adjacency matrix A of the graph, i.e., the structural information and the filtered attribute information as the input of the model, and use IAE and GCN to encode the input
[0030] S2.2: The autoencoder can be divided into an encoder and a decoder, and the encoding and decoding processes are shown in Equations 3 and 4
[0031]
[0032] where represents the output of the l-th layer encoder, represents the output of the l-th layer decoder, W and b represent the weight matrices and bias matrices of different layers, is the activation function is the input of the 0-th layer, take the filtered attribute matrix as the input, i.e., Improve the autoencoder through the context transfer mechanism, and the form is shown in Equation 5
[0033]
[0034] L is the number of layers of the encoder and decoder. The representation of the corresponding decoder is adjusted to the form shown in Equation 6
[0035]
[0036] By stacking L layers, the final generated representation is obtained
[0037] S2.3: The input of the GCN is the adjacency matrix of the graph and the attribute information. By encoding the attribute information and the structural information, the corresponding representation Z is generated, as shown in Equation 7.
[0038]
[0039] Among them, Z (l) represents the representation output by the l-th layer, and W (l) is the weight matrix of the corresponding layer.
[0040] Furthermore, in step S3:
[0041] S3.1: In order to obtain a more comprehensive and more suitable representation for downstream clustering tasks, the representations from IAE and GCN are dynamically fused, as shown in Equation 8.
[0042]
[0043] By this form, each layer of IAE and GCN is connected, and the structural information and attribute information of the graph are more fully aggregated, so that the finally generated representation retains more features and structures, avoiding the over-smoothing problem.
[0044] S3.2: Take as the input of the corresponding layer in the GCN to generate the final representation Z (L) , as shown in Equation 9.
[0045]
[0046] L is the number of stacked layers of the graph convolutional network.
[0047] Furthermore, in step S4.
[0048] S4.1: Take the finally generated representations H and Z of IAE and GCN as the input of the dual-supervision module, and unify IAE and GCN in a unified framework through the dual-supervision module.
[0049] S4.2: For the i-th node and the j-th category, use the t-distribution to measure the similarity between the data representation h i and the clustering center μ j , as shown in Equation 10.
[0050]
[0051] Among them, h i is the i-th row of jThe clustering is obtained by using the k-means algorithm on the representations learned by the pre-trained autoencoder. v is the degree of freedom of the t-distribution. Q = q ij can be regarded as the distribution of all samples.
[0052] S4.3: After obtaining the distribution of the samples, optimize the encoding process through the high-confidence distribution, as shown in Equation 11.
[0053]
[0054] where f j = ∑ j q ij represents the soft clustering frequency. In this way, the representation of each node is closer to the center of the data category, improving the clustering effect.
[0055] S4.4: Based on the obtained distributions P and Q, use the KL divergence to optimize the model training, as shown in Equation 12.
[0056]
[0057] S4.5: After multiple iterations of training, the method will obtain a stable representation, and use the soft assignment of the GCN-generated representation Z as the final clustering result, as shown in Equation 13.
[0058] r i = arg maxz ij (13)
[0059] Furthermore, in step S5:
[0060] S5.1: Use the out-of-bag method to divide the test set from the downloaded dataset and test the model with the test set to verify the effect of the method;
[0061] S5.2: Use the k-fold cross-validation method to test the effect of the method, where k = 10, randomly divide the test set into 10 parts to evaluate the effect of the model, and take the average of the 10 results;
[0062] S5.3: Combine and test the divided multiple test sets to evaluate the clustering ability of the method in practical applications;
[0063] S5.4: Compare the clustering results with the true categories to which the samples belong, and use the metrics ACC, NMI, ARI, and F1 to measure the effect of the model.
[0064] Other advantages, objects and features of the present invention will be set forth in the following description, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be learned from the practice of the present invention. The objects and other advantages of the present invention can be achieved and obtained through the following description. Description of the Drawings
[0065] In order to make the objects, technical solutions and beneficial effects of the present invention clearer, the present invention provides the following drawings for description:
[0066] Figure 1 It is the overall flowchart of the graph clustering method based on representation fusion proposed by the present invention. Detailed Embodiments
[0067] To make the inventive technical solutions, advantages and objects of the present invention clearer, the present invention will be further described below in conjunction with the drawings and specific embodiments.
[0068] As Figure 1 shown, the present invention proposes a graph clustering method based on representation fusion. The entire model includes 4 modules: a low-pass filter module, an encoding module, a representation fusion module, and a dual supervision module.
[0069] The low-pass filter module: constructs a parameterless low-pass filter based on the graph convolutional network, and uses the attribute information and structural information of the graph to filter the high-frequency noise in the attribute information. The filtered attribute information is more suitable for clustering.
[0070] The encoding module: uses an improved autoencoder (IAE) and a graph convolutional network (GCN) to encode the attribute information and structural information of the graph data, and generates corresponding embedded representations H and Z.
[0071] The representation fusion module: adopts a dynamic representation fusion mechanism to finely process the attribute and structural information extracted from the improved autoencoder (IAE) and the graph convolutional network (GCN), so as to obtain a more comprehensive and accurate representation structure.
[0072] The dual supervision module: optimizes the model training and clustering results by extracting high-confidence distributions, self-supervision and mutual-supervision to jointly supervise the clustering and encoding processes.
[0073] The technical solutions adopted by the present invention to achieve the above-mentioned invention objects are as follows:
[0074] S1: Data collection and processing: Select a public data set as the initial data, and perform low-pass filtering on the attribute information of the graph data through the low-pass filter module to filter high-frequency noise and generate filtered attribute information;
[0075] S2: Representation learning: Encode the attribute information and structural information through IAE and GCN to generate corresponding embedding representations;
[0076] S3: Representation fusion: Dynamically fuse the representations from IAE and GCN through a representation fusion module, so as to learn a more suitable representation for clustering, avoid the over-smoothing problem, and improve the clustering performance.
[0077] S4: Iterative training: Combine the generated representations above to optimize model training, and supervise the representation learning and clustering processes through a dual-supervision module to improve the clustering results.
[0078] S5: Effect evaluation: After training the model according to the training set, test the actual effect of the model with the test set, and evaluate the clustering ability of the model according to common model performance evaluation metrics.
[0079] S4: Iterative training: Combine the generated representations above to optimize model training, and supervise the representation learning and clustering processes through a dual-supervision module to improve the clustering results.
[0080] In the above step S1:
[0081] S1.1: Download common graph clustering datasets: Cite, Cora, ACM, Pubmed from the Cornell University Digital Library.
[0082] S1.2: The Cora dataset consists of seven different categories of scientific papers. It includes 2,708 papers, and each paper is represented as a node in the network. Cite is a citation network dataset consisting of papers from six different research categories: Agent, Artificial Intelligence (AI), Database (DB), Information Retrieval (IR), Machine Learning (ML), and Human-Computer Interaction (HCI). The ACM dataset is a network of paper relationships sourced from the ACM database. The Pubmed dataset consists of biomedical literature, with each literature regarded as a node and citation relationships regarded as edges. The detailed information of the four datasets is shown in Table 1:
[0083] Table 1 Dataset detailed information
[0084]
[0085] S1.3: The downloaded dataset file format is npy file. Write a script program to change it to a txt file for model training.
[0086] S1.4: The description of the dataset is as follows: Given an attributed graph G=(V, E, X), where is the vertex set, n is the number of nodes in the graph, E is the set of edges in the graph, and X =
[0087] [x1, x2, …, x n T is the attribute information of the graph. The topological structure of the graph can be represented by an adjacency matrix . If v ij = 1, it means there is an edge between node v i and node v j . Otherwise, it means there is no edge between the corresponding nodes. denotes the degree matrix of A, where denotes the degree of node v i .
[0088] S1.4: Construct a parameterless low-pass filter, and the form of the filter is shown in Equation 14.
[0089]
[0090] where F is the filter, I is the identity matrix, is the symmetric normalized Laplacian matrix, is the normalized adjacency matrix, is the normalized degree matrix.
[0091] S1.5: Perform low-pass filtering on the attribute information of the graph data, and the formula is shown in Equation 15.
[0092]
[0093] where t represents the stacking layer number of the filter, represents the attribute information after filtering high-frequency noise.
[0094] In the said step S2:[[]]
[0095] S2.1: Take the adjacency matrix A of the graph, i.e., the structure information, and the filtered attribute information as the input of the model, and use IAE and GCN to encode the input.
[0096] S2.2: The autoencoder can be divided into an encoder and a decoder. The encoding process of the encoder and the decoding process of the decoder are as
[0097] shown in Equations 16 and 17.
[0098]
[0099] where represents the output of the l-th layer encoder, represents the output of the l-th layer decoder, W and b represent the weight matrix and bias matrix of the encoder and decoder of the corresponding layer, is the activation function. For the input of the 0th layer, the filtered attribute matrix is used as the input, that is The decoder is mainly used to decode the representation output by the encoder and reconstruct the input based on the generated representation. In order to make the reconstructed input contain more attribute information, the autoencoder is improved based on the context passing mechanism, and the improvement method is shown in Equation 18.
[0100]
[0101] L is the number of layers of the encoder and decoder. The representation output by each layer of the encoder is input into the decoder through the context passing mechanism. The corresponding representation of the decoder is adjusted to the form shown in Equation 19.
[0102]
[0103] By stacking L layers, the final generated representation is obtained
[0104] S2.3: The input of the GCN is the adjacency matrix and attribute information of the graph, and the corresponding representation Z is generated by encoding the attribute information and structural information, as shown in Equation 20.
[0105]
[0106] Among them, Z (l) represents the representation output by the lth layer, and W (l) is the weight matrix of the corresponding layer.
[0107] In the step S3:
[0108] S3.1: Considering that the representation learned by the autoencoder can reconstruct the input data and contains rich attribute information, the representations from the IAE and the GCN are dynamically fused to obtain a more comprehensive and more suitable representation for downstream clustering tasks, as shown in Equation 21.
[0109]
[0110] By connecting each layer of the IAE and the GCN in this form, the structural information and attribute information of the graph are more fully aggregated, so that the finally generated representation retains more feature information and structural information, effectively avoiding the over-smoothing problem.
[0111] S3.2: Take as the input of the corresponding layer in the GCN to generate the final representation Z (L) , as shown in Equation 22.
[0112]
[0113] L is the number of stacked layers of the graph convolutional network. The last layer is a multi-classification layer with a softmax function as shown in Equation 23.
[0114]
[0115] The result z ij ∈Z represents the probability that sample i belongs to cluster center j, that is, the probability distribution. In addition, the input of the first-layer GCN is the filtered attribute information, as shown in Equation 24.
[0116]
[0117] In the said step S4:
[0118] S4.1: Use the representations H and Z finally generated by IAE and GCN as the input of the dual-supervision module, and unify IAE and GCN in a unified framework through the dual-supervision module.
[0119] S4.2: For the i-th node and the j-th category, use the t-distribution to measure the similarity between the data representation z i and the cluster center μ j as shown in Equation 25.
[0120]
[0121] where z i is the i-th row of Z, μ j is the cluster obtained by using the k-means algorithm on the representation learned by the pre-trained autoencoder. v is the degree of freedom of the t-distribution. Q = q ij can be regarded as the distribution of all samples.
[0122] S4.3: After obtaining the distribution of the samples, in order to improve the cluster distribution, optimize the encoding process through the high-confidence distribution, as shown in Equation 26.
[0123]
[0124] where f j = ∑ j q ij represents the soft cluster frequency. In this form, the representation of each node is closer to the center of the data category, improving the clustering effect.
[0125] S4.4: Based on the obtained distributions P and Q, use the KL divergence to optimize the model training, as shown in Equation 27.
[0126]
[0127] Since the distribution P is calculated from the distribution Q and the KL divergence loss between the two refines the clustering, this strategy is regarded as a self-supervised strategy.
[0128] S4.5: Similar to the self-supervised strategy, the distribution Q' is calculated based on the representation H generated by IAE, and then the KL divergence between P and Q' is calculated according to Equation 27, which is mutual supervision.
[0129] S4.5: After multiple iterations of training, the method will obtain a stable representation, and the soft assignment of the GCN-generated representation Z is used as the final clustering result, as shown in Equation 28.
[0130] r i =arg maxz ij . (28)
[0131] Furthermore, in step S5:
[0132] S5.1: Using the out-of-flow method, a test set is divided from the downloaded dataset, and the model is tested with the test set to verify the effectiveness of the method;
[0133] S5.2: Using the k-fold cross-validation method to test the effectiveness of the method, where k = 10, the test set is randomly divided into 10 parts to evaluate the effectiveness of the model, and the average of the 10 results is taken;
[0134] S5.4: Combine and test multiple divided test sets to evaluate the clustering ability of the method in practical applications;
[0135] S5.5: Compare the clustering results with the true categories to which the samples belong, and use the metrics ACC, NMI, ARI, and F1 to measure the effectiveness of the model.
[0136] ACC is the ratio of the number of correctly classified samples in the clustering result to the total number of samples. In clustering, it is usually necessary to match the clustering result with the true label and then calculate the accuracy rate, and the formula is as shown in 29.
[0137]
[0138] NMI, that is, the normalized mutual information, is the ratio of the mutual information (MI) to the average of the entropy of the true label and the clustering result. It measures the common information between two label assignments, and the calculation formula is as shown in 30.
[0139]
[0140] Among them, U is the true label, V is the clustering result, MI(U, V) is the mutual information, and H(U) and H(V) are the entropy of U and V respectively.
[0141] ARI, that is, the Adjusted Rand Index, is an adjusted version of the Rand Index (RI) that takes into account the randomness of label assignment. The Rand Index measures the agreement between two label assignments, and its calculation formula is as shown in 31.
[0142]
[0143] F1, that is, the F1 score, is the harmonic mean of recall and precision. In clustering, it is usually necessary to match the clustering results with the true labels and then calculate the F1 score, and the formula is as shown in 32.
[0144]
[0145] Among them, TP, FP, FN, and TN are the four core metrics for evaluating the performance of a classification model. TP, that is, True Positive, refers to the number of samples that the model correctly predicts as the positive class. That is, samples that are actually in the positive class and the model also predicts as the positive class. FP, that is, False Positive, refers to the number of samples that the model incorrectly predicts as the positive class. That is, samples that are actually in the negative class (normal), but the model predicts as the positive class. FN, that is, False Negative, refers to the number of samples that the model incorrectly predicts as the negative class. That is, samples that are actually in the positive class, but the model predicts as the negative class. TN, that is, True Negative, refers to the number of samples that the model correctly predicts as the negative class. That is, samples that are actually in the negative class (normal) and the model also predicts as the negative class.
[0146] As described above, it is only the specific implementation manner in the present invention, but the protection scope of the present invention is not limited thereto. Any transformation or replacement that can be understood and conceived by those familiar with the technology within the technical scope disclosed by the present invention should be covered within the scope of the present invention.
Claims
1. The present invention belongs to the field of graph clustering, and specifically relates to a graph clustering method based on representation fusion. The model includes a low-pass filter module, an encoding module, a representation fusion module and a dual supervision module. The low-pass filter module constructs a parameter-free low-pass filter based on the graph convolutional network, and uses the attribute information and structural information of the graph to filter out high-frequency noise in the attribute information. The filtered attribute information is more suitable for clustering. The encoding module uses an improved autoencoder (IAE) and a graph convolutional network (GCN) to encode the attribute information and structural information of the graph data and generate corresponding embedded representations H and Z. The representation fusion module adopts a dynamic representation fusion mechanism to finely process the attribute and structure information extracted from the improved autoencoder (IAE) and graph convolutional network (GCN), thereby obtaining a more comprehensive and accurate representation structure. The dual supervision module optimizes model training and clustering results by extracting high confidence distributions, self-supervision and mutual supervision to jointly supervise the clustering and encoding processes. It is characterized in that The following steps are involved: S1: Data collection and processing: Select the public data set as the initial data, perform low-pass filtering on the attribute information of the graph data through the low-pass filter module, filter out high-frequency noise, and generate filtered attribute information; S2: Representation learning: Encode attribute information and structural information through IAE and GCN to generate corresponding embedded representations; S3: Representation fusion: The representations from IAE and GCN are dynamically fused through the representation fusion module to learn representations that are more suitable for clustering, avoid over-smoothing problems, and improve clustering performance. S4: Iterative training: Combine the above generated representations to optimize model training, and use a dual supervision module to supervise the representation learning and clustering process to improve the clustering results. S5: Effect evaluation: After training the model based on the training set, use the test set to test the actual effect of the model, and evaluate the clustering ability of the model based on common model performance evaluation indicators.
2. The graph clustering method based on representation fusion according to claim 1, characterized in that: The following steps are involved: S1.1: Download commonly used graph clustering datasets from Cornell University Digital Library: Cite, Cora, ACM, Pubmed. S1.2: The downloaded dataset file format is npy file. Write a script to convert it into a txt file for model training. S1.3: The dataset is described as follows: Given an attribute graph G = (V, E, X), where v = {v1, v2, …, v n } is the vertex set, n is the number of nodes in the graph, E is the sum of the edges in the graph, X = [x1, x2, ..., x n ] T is the attribute information of the graph. The topological structure of the graph can be expressed using the adjacency matrix It means that if v ij =1, indicating node v i To node v j There is an edge between the corresponding nodes, otherwise there is no edge between the corresponding nodes. represents the degree matrix of A, where Represents node v i The degree. S1.4: Construct a low-pass filter without parameters. The form of the filter can be expressed as: Among them, F is the filter, I is the unit matrix, is the symmetric normalized Laplacian matrix, is the normalized adjacency matrix, is the normalized degree matrix. S1.5: Perform low-pass filtering on the attribute information of the graph data. The formula is as follows: Where t represents the number of filter stacking layers, Represents the attribute information after filtering high-frequency noise.
3. The graph clustering method based on representation fusion according to claim 1, characterized in that: The following steps are involved: S2.1: The adjacency matrix A of the graph, i.e., the structural information and the filtered attribute information As the input of the model, IAE and GCN are used to encode the input. S2.2: The autoencoder can be divided into an encoder and a decoder. The encoding and decoding process can be expressed as: in, represents the output of the encoder at layer l, represents the output of the l-th layer decoder, W and b represent the weight matrices and bias matrices of different layers, is the activation function. As the input of layer 0, we take the filtered attribute matrix as input, that is, The autoencoder is improved by the context transfer mechanism, as follows: L is the number of layers of the encoder and decoder. The corresponding decoder representation is adjusted to the following form: By stacking L layers, we get the final generated representation S2.3: The input of GCN is the adjacency matrix and attribute information of the graph. The corresponding representation Z is generated by encoding the attribute information and structure information in the following form: Among them, Z (l) represents the output of layer l, W (l) ) is the weight matrix of the corresponding layer.
4. The graph clustering method based on representation fusion according to claim 1, characterized in that: The following steps are involved: S3.1: In order to obtain a more comprehensive representation that is more suitable for downstream clustering tasks, we dynamically fuse the representations from IAE and GCN in the following form: By connecting each layer of IAE and GCN in this way, the structural and attribute information of the graph can be more fully aggregated, so that the final generated representation retains more features and structures and avoids the over-smoothing problem. S3.2: As the input of the corresponding layer in GCN, the final representation Z is generated (L) , of the following form: L is the number of stacked layers of the graph convolutional network.
5. The graph clustering method based on representation fusion according to claim 1, characterized in that: The following steps are involved: S4.1: The final representations H and Z generated by IAE and GCN are used as the input of the dual supervision module, which unifies IAE and GCN in a unified framework. S4.2: For the i-th node and the j-th category, use the t distribution to measure the data representation h i and cluster centers μ j The similarity between them is as follows: Among them, h i yes The i-th row of j is the clustering obtained by using the k-means algorithm on the representation learned by the pre-trained autoencoder. v is the degrees of freedom of the t distribution. Q = q ij can be considered as the distribution of all samples. S4.3: After obtaining the distribution of samples, we optimize the encoding process through high confidence distribution, as follows: Among them, f j =∑ j q ij Indicates the soft clustering frequency. In this form, the representation of each node is closer to the center of the data category, improving the clustering effect. S4.4: Based on the obtained distributions P and Q, KL divergence is used to optimize model training in the following form: S4.5: After multiple iterations of training, the method will obtain a stable representation, and the soft assignment of the GCN-generated representation Z is used as the final clustering result, which is in the following form: r i =arg maxz ij (13) 6. The graph clustering method based on representation fusion according to claim 1, characterized in that: The following steps are involved: S5.1: Use the outflow method to divide the test set from the downloaded data set, and use the test set to test the model to verify the effect of the method; S5.2: Use k-fold cross-validation to test the effectiveness of the method, where k = 10. The test set is randomly divided into 10 parts to evaluate the effectiveness of the model, and the 10 results are averaged. S5.3: Perform a combined test on the divided multiple test sets to evaluate the clustering ability of the method in practical applications; S5.4: Compare the clustering results with the true categories to which the samples belong, and use the indicators ACC, NMI, ARI and F1 to measure the effectiveness of the model.