Depth attribute graph clustering method based on dual information comparison strategy

Abstract

The invention relates to the field of depth attribute graph clustering, and particularly discloses a depth attribute graph clustering method based on a dual information comparison strategy, and the method comprises the steps: S1, providing a composite graph enhancement strategy which does not destroy original topological information, and generating a comparison view; s2, GAE and MLP are combined to design a linear graph automatic encoder, and information of the two views is transmitted to the linear graph automatic encoder for representation learning; s3, providing a dual information comparison strategy based on the comparison view obtained in S1, and performing comparison learning of the two views; s4, designing a multi-distribution self-supervision strategy, and forming a self-supervision cluster; s5, setting a multi-distribution self-supervision module to integrate the upstream task and the downstream task together, and performing global optimization; and S6, performing an ablation experiment to verify the effectiveness of the DAGC-DICS. According to the method, the clustering quality is remarkably improved, misclassification is reduced, and the clustering performance is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of deep attribute graph clustering, and specifically to a deep attribute graph clustering method based on a dual information comparison strategy. Background Art

[0002] The rapid development of information technology and the Internet has led to the emergence of various network data. In real life, there are also various network data, such as social networks, citation networks, and relationship networks. These networks contain not only attribute information, but also structural information. Therefore, these networks are naturally represented as a new type of graph, called attribute graph. One of the basic tasks of network analysis is node clustering. Node clustering has always been favored by researchers. With the continuous development of computer technology, the previous clustering method using only machine learning methods is no longer suitable for large-scale data sets. At present, machine learning is basically combined with deep learning for node clustering. Deep attribute graph clustering uses deep learning to cluster attribute graphs. By measuring the similarity between nodes, nodes are clustered into unconnected clusters, so that the similarity of nodes within a cluster is high and the similarity of nodes between clusters is low. Deep attribute graph clustering has many applications in real life, such as anomaly detection, community detection, and recommendation systems. However, deep attribute graphs still face many challenges at this stage. The most concerned issue among scholars is how to effectively integrate structural information and attribute information.

[0003] Deep attribute graph clustering is a basic task in graph learning. Existing deep attribute graph clustering is mainly divided into two stages: upstream graph representation learning and downstream clustering tasks. The upstream uses graph representation learning to obtain node embeddings, and the downstream uses the K-Means method to cluster the node embeddings obtained from the upstream learning.

[0004] In recent years, graph neural networks (GNNs) have developed rapidly, and methods for learning attribute graph representations based on GNNs have emerged one after another. Most of the previous studies are based on supervised learning or semi-supervised learning, but there is a big problem that label guidance is required in the process of graph representation learning. However, in practical applications, obtaining effective labels is a very challenging task and consumes a lot of manpower and material resources in the process. In order to circumvent this problem, at this stage, unsupervised methods or self-supervised methods are basically used for graph representation learning. Deep attribute graph clustering using this method can be roughly divided into two categories: attribute graph clustering based on reconstruction and attribute graph clustering based on contrastive learning. The reconstruction-based method uses GNN or its derivative network as an encoder to encode the structural information and attribute information of the input graph, and then reconstructs the obtained node embedding through the decoder to obtain new attribute information and structural information. Attribute graph clustering based on contrastive learning uses an enhancement strategy to obtain an augmented graph of the original graph before graph representation learning, and then performs graph representation learning on the original graph and augmented graph respectively. After obtaining the node embeddings of the two views, contrastive learning is performed to maximize the aggregation of positive sample pairs with high similarity and distance the negative sample pairs with low similarity. Both methods first obtain node embeddings through graph representation learning, and then use various clustering methods for clustering.

[0005] Although graph representation learning based on contrastive learning has received great attention and achieved good results, both of the above methods have a problem that they rely heavily on the representation learning ability of the network model. Even if node embedding is obtained through graph representation learning, it cannot be guaranteed that the optimal solution can be obtained by using it for clustering, because these methods are two-stage methods, and the upstream and downstream tasks are separate. In many cases, reconstruction-based methods are easily affected by reconstruction strategies and cannot fully utilize the potential information encoded in low-dimensional representations. At present, some scholars believe that although the reconstruction method is simple and easy to implement, its development and graph representation learning ability have lagged behind contrastive learning in recent years. With the widespread application of contrastive learning in image processing, target detection and other fields, it has become a trend to combine contrastive learning and graph representation learning. The key to contrastive learning is how to design effective enhancement strategies to construct contrastive views. For attribute graphs, the existing topological structure and node attribute random enhancement strategies may introduce noise or even destroy the structure of the original graph, thereby affecting the performance of downstream clustering. Summary of the invention

[0006] In view of the above-mentioned problems of the prior art, the present invention provides a deep attribute graph clustering method with a composite graph enhancement strategy and a dual information comparison strategy, which can effectively take into account both attribute information and structural information and perform comparative learning from different levels, thereby capturing more discriminative representations, enhancing subsequent clustering, and integrating upstream tasks and downstream tasks into an end-to-end framework, while optimizing upstream tasks and downstream tasks.

[0007] The technical solution adopted by the present invention is as follows:

[0008] A deep attribute graph clustering method based on a dual information comparison strategy includes the following steps:

[0009] S1. A composite graph enhancement strategy is proposed to generate contrast views without destroying the original topological information.

[0010] S2. Combine GAE and MLP to design a linear graph autoencoder, and pass the information of the two views into the linear graph autoencoder for representation learning;

[0011] S3, based on the contrast view obtained in S1, a dual information contrast strategy is proposed to conduct contrast learning on the two contrast views;

[0012] S4. Design a multi-distribution self-supervision strategy that uses the learned node representation to generate cluster assignments and combines the distribution consistency constraints of different views to form self-supervised clustering;

[0013] S5. Set up a multi-distributed self-supervision module to integrate the upstream task and the downstream task together, train the two tasks end-to-end at the same time by constructing a joint loss function, and then use the optimizer to globally optimize each module;

[0014] S6. Conduct ablation experiments to verify the effectiveness of DAGC-DICS.

[0015] Preferably, the method comprises four main mechanisms: a composite graph enhancement strategy, a linear graph autoencoder, a dual information comparison strategy, and a multi-distribution self-supervision module.

[0016] Preferably, in S1, the process of generating the comparison view is as follows:

[0017] S11, calculate the similarity matrix between nodes based on the attributes, and use the KNN method to generate a nearest neighbor graph based on node features;

[0018] S12. Calculate node v using cosine similarity i and v j The attribute vector x i and x j The similarity of ij , get the similarity matrix S∈R of all nodesN×N ;

[0019] S13. Select k pairs of similar nodes for each node obtained in S12 and connect them to obtain an undirected KNN attribute graph G. K = {V,E,X,A K}, where V = {v1,v2,........v N} is a set with N nodes and K classes, E represents the number of edges in the undirected graph, A K ∈R N×N is the reconstructed adjacency matrix, the attribute matrix X∈R N×d Remain unchanged, N represents the number of nodes, and d represents the feature dimension;

[0020] S14, introduce edge betweenness centrality to update and filter G K The connection relationship in the updated graph is represented by G′ K ={V,E,X,A′}, where A′ is the new adjacency matrix;

[0021] S15. Use Gaussian distribution to obtain a new feature matrix X′, G′={V, E, X′, A′} as a comparison view of the input.

[0022] Preferably, in S2, the linear graph autoencoder consists of multiple linear layers with biases and one graph neural network layer, and its formal expression is:

[0023]

[0024]

[0025] Among them, Ψ() is the activation function, is the normalized adjacency matrix obtained by renormalization technique, X represents the attribute matrix, and b l Represent the weight matrix and bias of the linear layer, ω g represents the weight matrix of GCN, z is the representation learned by GCN (graph neural network), Z (i+1) is the representation obtained after learning the i+1th linear layer, which serves as the input of the subsequent GCN. is the bias of the i+1th linear layer.

[0026] Preferably, the attribute matrix X of the original data is passed into the first layer of the linear graph autoencoder to obtain the representation Z (1) , whose expression is:

[0027]

[0028] Preferably, in S3, the dual information comparison strategy mainly includes the comparison learning of node embedding and reconstructed adjacency matrix.

[0029] Preferably, in S3, the specific process of the dual information comparison strategy for comparative learning of two contrasting views is as follows:

[0030] S31, respectively pass the two views through the linear graph autoencoder with shared parameters in S2 to perform graph representation learning, obtain node embeddings of the two views, and construct positive and negative sample pairs for comparative learning;

[0031] S32. After obtaining the node embeddings of the two views, the adjacency matrices of the two views are reconstructed by calculating the inner product, and comparative learning is performed on the reconstructed adjacency matrices of the two views.

[0032] Preferably, in S3, the total loss of the dual information contrast strategy includes the node embedding contrast loss and the reconstructed adjacency matrix contrast loss, and the specific form is as follows:

[0033] L=λ1L Node +λ2L adj ;

[0034] Where λ1 and λ2 are the importance weights of balancing node embedding comparison and reconstructing adjacency matrix comparison, respectively, and L Node is the node embedding contrast loss, L adj Contrastive loss for reconstructing the adjacency matrix;

[0035] Among them, InfoNCE loss is used to represent the contrast loss of node embedding, which is expressed as follows:

[0036]

[0037] Where sim() is the cosine similarity, t is the temperature parameter, (z i ,z i ′) is a positive sample pair, z i The node encoding representation corresponding to the original view, z i ′ is the node encoding representation of the contrast view;

[0038] The mean square error loss MSE is used to represent the contrast loss of reconstructing the adjacency matrix. The inner product solution and MSE loss are shown as follows:

[0039]

[0040] Preferably, in S4, the specific steps of adopting the multi-distribution self-supervision strategy to form self-supervision clustering are as follows:

[0041] S41. Calculate the cluster center point c j and node representation zi The similarity between them is converted into a t-distribution that conforms to Students, where i∈{1,2,....,N} and j∈{1,2,.....K}. K-Means is used to initialize and randomly restart on the pre-trained node representation, and an internal clustering effective indicator is used to select the best solution;

[0042] S42, using Students' t distribution to cluster the center point c j and node representation z i The similarity between them is converted into probability distribution, and the obtained soft clustering assignment Q = [q ij ]∈R N×K , where the cluster prediction q ij Assign node i to centroid c j The soft clustering assignment probability of is expressed as:

[0043] S43, normalize Q square to generate an enhanced clustering target distribution P = [p ij ]∈R N×K , the network is trained as a high confidence guide, and its calculation formula is as follows: Among them, m∈{1,2,....K}, the total frequency f j =∑ j q ij Indicates the assignment to cluster center c j The sum of the probabilities of

[0044] S44. By minimizing the KL distance between Q and P, the cluster distribution Q is made closer to the target distribution P, the learned representation is closer to the cluster center, and the cluster soft assignment is refined, which is defined as: Among them, L pq1 Update the representation learned by LGAE in a self-supervised manner;

[0045] S45. Use the enhanced representation H of the node embedding of the original view to generate the soft cluster assignment Q. Learn the soft cluster assignment Q′ by comparing the potential representation Z′ obtained from the view. Minimize the KL divergence between Q′ and P to supervise the distribution of Q′, which is defined as:

[0046]

[0047] S46, by minimizing the KL divergence between Q and Q′, the two soft clustering assignments learn from each other and optimize the LGAE extraction features, then the objective function is defined as:

[0048]

[0049] Preferably, in S5, the specific steps of global optimization are as follows:

[0050] S51. The network learning loss is divided into two parts: the loss of contrastive learning and the clustering loss of the multi-distribution self-supervision module, which is expressed as:

[0051] S52. During the network training process, the Adam optimizer is used to perform momentum optimization on the network parameters and cluster centers of each module;

[0052] S53. After the network training is completed, the total loss of the model converges, and the output of the model consists of the node embeddings of the two views. The soft clustering assignments Q and Q′ of the node embeddings of the two views are obtained respectively.

[0053] S54, select Q as the clustering result, and then determine the predicted label of the cluster according to Q. The calculation formula is as follows:

[0054] The present invention has the following characteristics and advantages:

[0055] (1) A new comparison view that does not destroy the original topological structure is constructed, and a dual information comparison strategy is designed to promote the aggregation of similar nodes and the separation of dissimilar nodes;

[0056] (2) Effectively alleviate the representation collapse of graph representation learning, taking into account both attribute information and structural information when performing contrastive learning, thereby capturing more discriminative representations;

[0057] (3) Integrate upstream and downstream tasks into an end-to-end framework and optimize both upstream and downstream tasks to ensure that the representation learned by the network is suitable for clustering tasks;

[0058] (4) We take advantage of the advantages of graph contrastive learning and deep graph clustering and gain in-depth understanding of the design of unified deep attribute graph clustering. Compared with the most advanced methods, our model has certain advantages in clustering.

[0059] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments, but the embodiments should not be construed as limiting the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is the overall framework of DAGC-DICS proposed in the present invention;

[0061] Figure 2 It is an ablation study of three benchmarks on four metrics. DETAILED DESCRIPTION

[0062] In order to make the technical solutions, advantages and purposes of the present invention clearer, the technical solutions of the embodiments of the present invention are clearly and completely described below. The described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of this application.

[0063] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0064] like Figure 1 As shown in the figure, the proposed DAGC-DICS overall framework contains four main mechanisms: composite graph enhancement strategy, linear graph autoencoder, dual information comparison strategy and multi-distribution self-supervision module. DAGC-DICS integrates upstream and downstream tasks into an end-to-end framework for global joint optimization. The entire model is roughly divided into three parts. First, a comparative view is generated through a composite graph enhancement strategy that does not destroy the original topological information. Then, the information of the two views is passed into the linear graph autoencoder for representation learning. Then, the node embeddings learned from the two views are respectively calculated by the inner product to obtain the reconstructed adjacency matrix. The obtained adjacency matrix and node embedding are simultaneously input into the dual information comparison module for comparative learning. The learned node representation is used to generate cluster assignments, and the distribution consistency constraints of different views are combined to form self-supervised clustering.

[0065] The present invention proposes a deep attribute graph clustering method based on a dual information comparison strategy, comprising the following steps:

[0066] S1. A composite graph enhancement strategy is proposed that does not destroy the original topological information to generate a contrast view. The process of generating a contrast view is as follows:

[0067] S11, calculate the similarity matrix between nodes based on the attributes, and use the KNN method to generate a nearest neighbor graph based on node features;

[0068] S12. Calculate node v using cosine similarity i and v j The attribute vector x i and x j The similarity of ij , get the similarity matrix S∈R of all nodes N×N ;

[0069] S13. Select k pairs of similar nodes for each node obtained in S12 and connect them to obtain an undirected KNN attribute graph G. K = {V,E,X,A K}, where V = {v1,v2,........v N} is a set with N nodes and K classes, E represents the number of edges in the undirected graph, A K ∈R N×N is the reconstructed adjacency matrix, the attribute matrix X∈R N×d Remain unchanged, N represents the number of nodes, and d represents the feature dimension;

[0070] S14, introduce edge betweenness centrality to update and filter G K The connection relationship in the updated graph is represented by G′ K ={V,E,X,A′}, where A′ is the new adjacency matrix;

[0071] S15. Use Gaussian distribution to obtain a new feature matrix X′, G′={V, E, X′, A′} as a comparison view of the input.

[0072] S2. Combine GAE and MLP to design a linear graph autoencoder, and pass the information of the two views into the linear graph autoencoder for representation learning;

[0073] In S2, the linear graph autoencoder consists of multiple linear layers with biases and a layer of graph neural network. Its formal expression is:

[0074]

[0075] Among them, Ψ() is the activation function, is the normalized adjacency matrix obtained by renormalization technique, X represents the attribute matrix, and b l Represent the weight matrix and bias of the linear layer, ω g represents the weight matrix of GCN, z is the representation learned by GCN (graph neural network), Z (i+1) is the representation obtained after learning the i+1th linear layer, which serves as the input of the subsequent GCN. is the bias of the i+1th linear layer.

[0076] In addition, the attribute matrix X of the original data is passed into the first layer of the linear graph autoencoder to obtain the representation Z (1) , whose expression is:

[0077]

[0078] S3, based on the contrast view obtained in S1, a dual information contrast strategy is proposed to conduct contrast learning on the two contrast views;

[0079] In S3, the dual information comparison strategy mainly includes the comparative learning of node embedding and reconstructed adjacency matrix;

[0080] In S3, the specific process of the dual information comparison strategy for contrastive learning of two contrasting views is as follows:

[0081] S31, respectively pass the two views through the linear graph autoencoder with shared parameters in S2 to perform graph representation learning, obtain node embeddings of the two views, and construct positive and negative sample pairs for comparative learning;

[0082] S32. After obtaining the node embeddings of the two views, the adjacency matrices of the two views are reconstructed by calculating the inner product, and comparative learning is performed on the reconstructed adjacency matrices of the two views.

[0083] In S3, the total loss of the dual information contrast strategy includes the node embedding contrast loss and the reconstructed adjacency matrix contrast loss, which is as follows:

[0084] L=λ1L Node +λ2L adj ;

[0085] Where λ1 and λ2 are the importance weights of balancing node embedding comparison and reconstructing adjacency matrix comparison, respectively, and L Node is the node embedding contrast loss, L adj Contrastive loss for reconstructing the adjacency matrix;

[0086] Among them, InfoNCE loss is used to represent the contrast loss of node embedding, which is expressed as follows:

[0087]

[0088] Where sim() is the cosine similarity, t is the temperature parameter, (z i ,z i ′) is a positive sample pair, z i The node encoding representation corresponding to the original view, z i ′ is the node encoding representation of the contrast view;

[0089] The mean square error loss MSE is used to represent the contrast loss of reconstructing the adjacency matrix. The inner product solution and MSE loss are shown as follows:

[0090]

[0091] S4. Design a multi-distribution self-supervision strategy that uses the learned node representation to generate cluster assignments and combines the distribution consistency constraints of different views to form self-supervised clustering;

[0092] In S4, the specific steps of using multi-distribution self-supervision strategy to form self-supervised clustering are as follows:

[0093] S41. Calculate the cluster center point c j and node representation z i The similarity between them is converted into a t-distribution that conforms to Students, where i∈{1,2,....,N} and j∈{1,2,.....K}. K-Means is used to initialize and randomly restart on the pre-trained node representation, and an internal clustering effective indicator is used to select the best solution;

[0094] S42, using Students' t distribution to cluster the center point c j and node representation z i The similarity between them is converted into probability distribution, and the obtained soft clustering assignment Q = [q ij ]∈R N×K , where the cluster prediction q ij Assign node i to centroid c j The soft clustering assignment probability of is expressed as:

[0095] S43, normalize Q square to generate an enhanced clustering target distribution P = [p ij ]∈R N×K , the network is trained as a high confidence guide, and its calculation formula is as follows: Among them, m∈{1,2,....K}, the total frequency f j =∑ j q ij Indicates the assignment to cluster center c j The sum of the probabilities of

[0096] S44. By minimizing the KL distance between Q and P, the cluster distribution Q is made closer to the target distribution P, the learned representation is closer to the cluster center, and the cluster soft assignment is refined, which is defined as: Among them, L pq1 Update the representation learned by LGAE in a self-supervised manner;

[0097] S45. Use the enhanced representation H of the node embedding of the original view to generate the soft cluster assignment Q. Learn the soft cluster assignment Q′ by comparing the potential representation Z′ obtained from the view. Minimize the KL divergence between Q′ and P to supervise the distribution of Q′, which is defined as:

[0098]

[0099] S46, by minimizing the KL divergence between Q and Q′, the two soft clustering assignments learn from each other and optimize the LGAE extraction features, then the objective function is defined as:

[0100]

[0101] S5. Set up a multi-distributed self-supervision module to integrate the upstream task and the downstream task together, train the two tasks end-to-end at the same time by constructing a joint loss function, and then use the optimizer to globally optimize each module;

[0102] In S5, the specific steps of global optimization are as follows:

[0103] S51. The network learning loss is divided into two parts: the loss of contrastive learning and the clustering loss of the multi-distribution self-supervision module, which is expressed as:

[0104] S52. During the network training process, the Adam optimizer is used to perform momentum optimization on the network parameters and cluster centers of each module;

[0105] S53. After the network training is completed, the total loss of the model converges, and the output of the model consists of the node embeddings of the two views. The soft clustering assignments Q and Q′ of the node embeddings of the two views are obtained respectively.

[0106] S54, select Q as the clustering result, and then determine the predicted label of the cluster according to Q. The calculation formula is as follows:

[0107] S6. Perform ablation experiments to verify the effectiveness of DAGC-DICS. The experimental process is as follows:

[0108] (1) Selection of data sets and clustering indicators

[0109] The datasets used are all authoritative public datasets in the field of deep graph clustering, and a large number of experiments are conducted on Citeseer, Cora, and Pubmed. The specific information of the dataset is shown in Table 1. We use four clustering indicators to evaluate the effectiveness of the proposed network. The four clustering indicators are: accuracy (ACC), normalized mutual information (NMI), average Rand index (ARI) and macro F1-score (F1). For these four clustering indicators, the larger their values, the better the performance.

[0110] Table 1: Dataset description

[0111] Dataset #Nodes #Edges #Dimension #Classes Citeseer 3327 4732 3703 6 Cora 2708 5429 1433 7 Pubmed 19717 44338 500 3

[0112] (2) Baseline selection

[0113] In order to better evaluate the clustering performance of our proposed DAGC-DICS on the above three datasets, we selected 10 basic methods as experimental comparison objects.

[0114] ARGA: It proposes a generative adversarial framework to regularize embeddings with prior distributions.

[0115] SDCN: It combines the attribute information learned by AE with that learned by GCN through the transfer operator to enrich the node embedding. It is an end-to-end graph clustering method.

[0116] MVGRL: A graph representation learning method based on multi-view contrastive learning is proposed. This method captures graph information of different scales through multiple different views.

[0117] GRACE: Corruptions are applied to the original graph to generate two views and maximize the consistency between node representations at different angles.

[0118] CCGC. It adopts two views with non-shared encoders and selects positive samples based on high-confidence clustered pseudo-labels.

[0119] CONVERT: A contrastive graph clustering network with robust enhancement is proposed, which employs a perturbation network and a restoration network for contrastive learning.

[0120] CLAGC: A multi-level contrastive mechanism based on pseudo-label guided negative sampling is proposed to maximize the consistency between node representations in multiple latent spaces.

[0121] (3) Experimental details

[0122] The overall architecture of LGAE consists of two linear layers and one graph neural network. Before formal training, pre-training was performed, with epochs set to 30, learning rate set to 0.001, and batch size set to 256. Then the entire network was trained with 200 epochs and optimized using the Adam optimizer. The selection of hyperparameters was done using a network search method: the dimension of LGAE was searched in {128, 256, 512}, the learning rate was searched in {0.00001, 0.00003, 0.00005, 0.0001, 0.0005}, the k of the KNN graph was adjusted in {1, 2, 3, 5, 9, 10}, and the value of λ was searched in {0.025, 0.05, 0.1, 0.2, 0.4, 0.8}. In order to initialize the cluster centroid, the K-Means method was run 20 times based on the representation obtained by pre-training, and the best result was selected. In order to prevent the clustering results from being accidental, the model was run 10 times, and its mean and standard deviation were taken to evaluate the clustering results to ensure the effectiveness of the clustering results.

[0123] The learning rates of each data set are as follows: For Cora, the learning rate is set to 2e -5 For Citeseer, the learning rate is set to 4e -5 For Pubmed, the learning rate is set to 2e -3For some baseline methods, the results in CONVERT and CLAGC are directly cited. All experiments are implemented on a server with 12vCPU Intel(R)Xeon(R)Platinum 8255C CPU@2.50GHz and NVIDIA GTX 3090 and 24GB memory.

[0124] (4) Performance comparison

[0125] Table 2 shows the clustering results of DAGC-DICS on four data sets. By observing Table 2, it can be found that the clustering indicators of the method proposed in the present invention on the three data sets are higher than those of the existing methods. This shows that the clustering results obtained by the method proposed in the present invention are consistent with the true labels. Especially for the PubMed data set, the four indicators of ACC, NMI, ARI and F1 index in CLAGC in the method proposed in the present invention are improved by 1.69, 0.54, 0.78 and 1.94 respectively. The improvement of ACC and NMI indicates that the overall performance of the model has been improved. The significant improvement of ARI and F1 index shows that our method has significantly improved the clustering quality of unbalanced data or certain categories, and reduced the misclassification between certain categories, thus further proving the effectiveness of the method proposed in the present invention.

[0126] Table 2: Clustering results on Cora

[0127] Method ACC NMI ARI F1 ARGA 71.04±0.25 51.06±0.52 47.71±0.33 69.27±0.39 SDCN 35.60±2.83 14.28±1.91 7.78±3.24 24.37±1.04 MVGRL 70.47±3.70 55.57±1.54 48.70±3.94 67.15±1.86 GRACE 61.30±3.81 47.10±3.06 38.73±4.65 58.06±3.43 CCGC 73.73±1.81 55.93±1.58 51.52±2.44 70.83±2.92 CONVERT 74.07±1.51 55.57±1.12 50.58±2.01 72.92±3.27 CLAGC 75.08±0.04 59.21±0.10 56.59±0.07 73.00±0.05 Our 75.99±0.21 60.51±0.11 58.03±0.32 74.02±0.13

[0128] Table 3: Clustering results on Citeseer

[0129]

[0130]

[0131] Table 4: Clustering results on Pubmed (OOM means out of memory)

[0132]

[0133] (5) Ablation experiment

[0134] Comprehensive ablation experiments are conducted to verify the effectiveness of the proposed module.

[0135] Four variants were designed for ablation analysis on Cora, CiteSeer, and PubMed datasets to study the advantages of the composite graph enhancement strategy (CGES), linear graph autoencoder (LGAE), dual information contrast strategy (DICS), and multi-distribution self-supervisory module (DDSM) modules. *CGES, *LGAE, *DICS, and *DDSM represent the network without composite graph enhancement strategy, the network without linear graph autoencoder, the network without dual information contrast strategy, and the network without multi-distribution self-supervisory module, respectively.

[0136] Specifically, when conducting ablation experiments, we use the strategy of attribute perturbation and random edge deletion instead of CGES, GAE with three-layer GCN instead of LGAE, and contrastive learning using only attribute information instead of DICS. For the DDSM module, we use the KL divergence of the original graph instead.

[0137] First observe Figure 2 It can be found that the complete network model is better than the model with missing modules in the four clustering indicators of the three data sets, showing its superiority. In particular, on the Cora and Citeseer data sets, we can clearly find that the clustering effect of the complete network model is much better than that of the model with missing modules.

[0138] From the above experiments, we can conclude that no matter which one of them has an impact on the clustering results, the lack of any module cannot make the clustering results optimal.

[0139] Therefore, the present invention proposes a deep attribute graph clustering method based on a dual information comparison strategy, which utilizes the advantages of graph contrast learning and deep graph clustering, has an in-depth understanding of the design of unified deep attribute graph clustering, effectively avoids the representation collapse problem in the graph representation learning process, integrates upstream tasks and downstream tasks into a framework, and optimizes the clustering loss and network loss together to ensure that the representation learned by the network is suitable for clustering tasks, significantly improves the clustering quality, reduces misclassification, and effectively improves the clustering performance.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A deep attribute graph clustering method based on a dual information comparison strategy, characterized in that: The steps include: S1. A composite graph enhancement strategy is proposed to generate contrast views without destroying the original topological information. S2. Combine GAE and MLP to design a linear graph autoencoder, and pass the information of the two views into the linear graph autoencoder for representation learning; S3, based on the contrast view obtained in S1, a dual information contrast strategy is proposed to conduct contrast learning on the two contrast views; S4. Design a multi-distribution self-supervision strategy that uses the learned node representation to generate cluster assignments and combines the distribution consistency constraints of different views to form self-supervised clustering; S5. Set up a multi-distributed self-supervision module to integrate the upstream task and the downstream task together, train the two tasks end-to-end at the same time by constructing a joint loss function, and then use the optimizer to globally optimize each module; S6. Conduct ablation experiments to verify the effectiveness of DAGC-DICS.

2. According to claim 1, a deep attribute graph clustering method based on a dual information comparison strategy is characterized in that: The method includes four main mechanisms: composite graph enhancement strategy, linear graph autoencoder, dual information comparison strategy and multi-distribution self-supervision module.

3. According to claim 1, a deep attribute graph clustering method based on a dual information comparison strategy is characterized in that: In S1, the process of generating the contrast view is as follows: S11, calculate the similarity matrix between nodes based on the attributes, and use the KNN method to generate a nearest neighbor graph based on node features; S12. Calculate node v using cosine similarity i and v j The attribute vector x i and x j The similarity of ij , get the similarity matrix S∈R of all nodes N×N ; S13. Select k pairs of similar nodes for each node obtained in S12 and connect them to obtain an undirected KNN attribute graph G. K = {V,E,X,A K }, where V = {v1,v2,........v N } is a set with N nodes and K classes, E represents the number of edges in the undirected graph, A K ∈R N ×N is the reconstructed adjacency matrix, the attribute matrix X∈R N×d Remain unchanged, N represents the number of nodes, and d represents the feature dimension; S14, introduce edge betweenness centrality to update and filter G K The connection relationship in the updated graph is represented by G′ K ={V,E,X,A′}, where A′ is the new adjacency matrix; S15. Use Gaussian distribution to obtain a new feature matrix X′, G′={V, E, X′, A′} as a comparison view of the input.

4. According to claim 1, a deep attribute graph clustering method based on a dual information comparison strategy is characterized in that: In S2, the linear graph autoencoder consists of multiple linear layers with biases and a layer of graph neural network. Its formal expression is: Among them, Ψ() is the activation function, is the normalized adjacency matrix obtained by renormalization technique, X represents the attribute matrix, and b l Represent the weight matrix and bias of the linear layer, ω g represents the weight matrix of GCN, z is the representation learned by GCN (graph neural network), Z (i+1) is the representation obtained after learning the i+1th linear layer, which serves as the input of the subsequent GCN. is the bias of the i+1th linear layer.

5. According to claim 4, a deep attribute graph clustering method based on a dual information comparison strategy is characterized in that: The attribute matrix X of the original data is passed into the first layer of the linear graph autoencoder to obtain the representation Z (1) , whose expression is:

6. According to claim 1, a deep attribute graph clustering method based on a dual information comparison strategy is characterized in that: In S3, the dual information comparison strategy mainly includes the comparative learning of node embedding and reconstructed adjacency matrix.

7. The deep attribute graph clustering method based on dual information comparison strategy according to claim 1 is characterized in that: In S3, the specific process of the dual information comparison strategy for contrastive learning of two contrasting views is as follows: S31, respectively pass the two views through the linear graph autoencoder with shared parameters in S2 to perform graph representation learning, obtain node embeddings of the two views, and construct positive and negative sample pairs for comparative learning; S32. After obtaining the node embeddings of the two views, the adjacency matrices of the two views are reconstructed by calculating the inner product, and comparative learning is performed on the reconstructed adjacency matrices of the two views.

8. The deep attribute graph clustering method based on dual information comparison strategy according to claim 7 is characterized in that: In S3, the total loss of the dual information contrast strategy includes the node embedding contrast loss and the reconstructed adjacency matrix contrast loss, which is as follows: L=λ1L Node +λ2L adj ; Where λ1 and λ2 are the importance weights of balancing node embedding comparison and reconstructing adjacency matrix comparison, respectively, and L Node is the node embedding contrast loss, L adj Contrastive loss for reconstructing the adjacency matrix; Among them, InfoNCE loss is used to represent the contrast loss of node embedding, which is expressed as follows: Where sim() is the cosine similarity, t is the temperature parameter, (z i ,z i ′) is a positive sample pair, z i The node encoding representation corresponding to the original view, z i ′ is the node encoding representation of the contrast view; The mean square error loss MSE is used to represent the contrast loss of reconstructing the adjacency matrix. The inner product solution and MSE loss are shown as follows:

9. The deep attribute graph clustering method based on dual information comparison strategy according to claim 1, characterized in that: In S4, the specific steps of using multi-distribution self-supervision strategy to form self-supervised clustering are as follows: S41. Calculate the cluster center point c j and node representation z i The similarity between them is converted into a t-distribution that conforms to Students, where i∈{1,2,....,N} and j∈{1,2,.....K}. K-Means is used to initialize and randomly restart on the pre-trained node representation, and an internal clustering effective indicator is used to select the best solution; S42, using Students' t distribution to cluster the center point c j and node representation z i The similarity between them is converted into probability distribution, and the obtained soft clustering assignment Q = [q ij ]∈R N×K , where the cluster prediction q ij Assign node i to centroid c j The soft clustering assignment probability of is expressed as: S43, normalize Q square to generate an enhanced clustering target distribution P = [p ij ]∈R N×K , the network is trained as a high confidence guide, and its calculation formula is as follows: Among them, m∈{1,2,....K}, the total frequency f j =∑ j q ij Indicates the assignment to cluster center c j The sum of the probabilities of S44. By minimizing the KL distance between Q and P, the cluster distribution Q is made closer to the target distribution P, the learned representation is closer to the cluster center, and the cluster soft assignment is refined, which is defined as: in, Update the representation learned by LGAE in a self-supervised manner; S45. Use the enhanced representation H of the node embedding of the original view to generate the soft cluster assignment Q. Learn the soft cluster assignment Q′ by comparing the potential representation Z′ obtained from the view. Minimize the KL divergence between Q′ and P to supervise the distribution of Q′, which is defined as: S46, by minimizing the KL divergence between Q and Q′, the two soft clustering assignments learn from each other and optimize the LGAE extraction features, then the objective function is defined as:

10. The deep attribute graph clustering method based on dual information comparison strategy according to claim 1, characterized in that: In S5, the specific steps of global optimization are as follows: S51. The network learning loss is divided into two parts: the loss of contrastive learning and the clustering loss of the multi-distribution self-supervision module, which is expressed as: S52. During the network training process, the Adam optimizer is used to perform momentum optimization on the network parameters and cluster centers of each module; S53. After the network training is completed, the total loss of the model converges, and the output of the model consists of the node embeddings of the two views. The soft clustering assignments Q and Q′ of the node embeddings of the two views are obtained respectively. S54, select Q as the clustering result, and then determine the predicted label of the cluster according to Q. The calculation formula is as follows:

Images