Method and system for multi-view clustering based on graph attention autoencoder

Abstract

The application provides a multi-view clustering method and system based on a graph attention automatic encoder, relates to the technical field of multi-view clustering, and specifically includes the following steps: selecting a view with the largest information quantity from different views of the same group of nodes; learning a graph structure and node content by using a trained graph attention encoder based on the view with the largest information quantity and node content information, so as to obtain a node feature representation; performing specific constraint on the node feature representation by using an l1,2-norm penalty, so as to obtain a constrained node feature representation; inputting the constrained node feature representation into a self-optimizing clustering module to perform clustering, so as to obtain a final clustering result; and the application applies the graph attention network to multi-view graph clustering, simultaneously reconstructs the graph structure and the node content, and makes the latent representation well preserve the graph structure and the content information of the nodes, so that the application is more suitable for cluster tasks.

Description

Technical Field

[0001] This invention belongs to the field of multi-view clustering technology, and particularly relates to a multi-view clustering method and system based on graph attention autoencoders. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In recent years, multi-view graph clustering has been increasingly proposed for analyzing and processing multimedia data. Most multi-view clustering algorithms typically involve two steps: first, constructing a shared similarity graph from the multi-view data; and then clustering this similarity graph to obtain the clustering results. In practical multimedia applications, due to the heterogeneity of multimedia acquisition sources, multi-view data often exhibits characteristics such as redundancy, correlation, and diversity. This leads to several problems with existing technologies:

[0004] 1) Most clustering methods use shallow models to handle the complex relationships within multi-view graphs, which severely limits the ability to model information from multi-view graphs;

[0005] The purpose of graph embedding is to learn a low-dimensional node representation while preserving the node's content information and topological structure. In the past decade, graph embedding methods have emerged. Based on different input information, graph embedding methods can be roughly divided into two categories: topology embedding methods (TSE) and content-enhanced graph embedding methods (CEGE).

[0006] The TSE method only takes the topology as input and maps it to learn low-dimensional node representations. For example, Perozzi et al. proposed a truncated random walk algorithm to learn node representations, which transforms the original graph structure information into a set of linear sequences. Unlike generating linear sequences, Cao et al. proposed a deep neural network for learning graph representations (DNGR), which proposes a random surfing model that can directly utilize topology information. Cavalari et al. integrated community embedding, community detection, and node embedding into a closed loop instead of embedding each node individually, proposing the Community Embedding Framework (ComE). To handle the problem of unknown community numbers, Cavalari et al. proposed learning finite and infinite community embedding graphs (ComE+). Although the above methods have achieved good results, they only consider the graph structure, which limits their performance. Studies have shown that multi-view data can help improve clustering performance; however, all the above methods only utilize the graph structure and node content of a single view, resulting in unsatisfactory results.

[0007] Numerous multi-view clustering methods are dedicated to learning high-quality latent representations or affinity matrices shared by different views. Among them, deep multi-view clustering methods have attracted widespread attention from researchers due to their outstanding representativeness and fast inference speed. For example, Andrew et al. proposed a new multi-view clustering algorithm that combines a deep encoder with canonical correlation analysis (DCCA). In order to better learn multi-view representations, Wang et al. extended DCCA by introducing a deep decoder and proposed the Deep Canonical Correlation Autoencoder (DCCAE).

[0008] To better utilize the graph structure information in multiple views, Li et al. proposed a multi-view learning method based on GCN, namely Co-GCN. However, Co-GCN is designed for semi-supervised clustering. To address this issue, Fan et al. proposed a multi-graph autoencoder (O2MAC) for graph embedding clustering. Although O2MAC is successful, it only encodes the node content information of a single view, resulting in limited performance when processing data of single-view graphs and node content.

[0009] 2) In recent years, many clustering methods for embedded graphs have been developed, but none of them consider the community-specific distribution of node representations, resulting in unsatisfactory clustering performance;

[0010] To date, several proposed clustering models are based on learning low-dimensional, compact, and continuous representations, and then applying classical clustering methods to these learned representations to obtain cluster labels. While this improves clustering performance, it ignores the cluster-specific distribution of node representations. Different communities are distributed across different feature dimensions, resulting in a highly chaotic distribution of node features. Even across most dimensions, node features are very similar, which may cause the algorithm to cluster them all into the same community, leading to low clustering results. Some of these methods only utilize the graph structure and node content of a single view, resulting in unsatisfactory results. Other methods utilize multiple views to process graph structure information, but only encode the node content information of a single view; their performance is limited when processing data with only single-view graphs and node content.

[0011] Therefore, how to integrate information from multiple views has become a topic worthy of research. Summary of the Invention

[0012] To overcome the shortcomings of the prior art, this invention provides a multi-view clustering method and system based on graph attention autoencoders. To better suit clustering tasks, graph attention networks are applied to multi-view graph clustering, while reconstructing the graph structure and node content, so that the latent representation can well preserve the graph structure and content information of the nodes.

[0013] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0014] The first aspect of this invention provides a multi-view clustering method based on a graph attention autoencoder;

[0015] Multi-view clustering methods based on graph attention autoencoders include:

[0016] Select the view with the most information from different views of the same set of nodes;

[0017] Based on the view and node content information with the most information, a trained graph attention encoder is used to learn the graph structure and node content to obtain node feature representations.

[0018] use - Norm penalty applies specific constraints to the node feature representation, resulting in a constrained node feature representation;

[0019] The constrained node feature representations are input into the self-optimizing clustering module for clustering to obtain the final clustering results;

[0020] Specifically, by adding a multi-view decoder, the constrained node feature representation is reconstructed, and the reconstruction loss is used to train the graph attention encoder.

[0021] Furthermore, the view with the most information is selected based on clustering metrics and adjacency matrices, by calculating the modular score of each graph view and selecting the graph view with the highest score as the view with the most information.

[0022] Furthermore, the graph attention encoder also learns the importance of neighboring nodes.

[0023] Furthermore, the learning of the importance of neighboring nodes is achieved by assigning different weights to neighbors in the layer-by-layer graph attention strategy of the graph attention encoder, representing the importance of neighboring nodes to the current node.

[0024] Furthermore, the aforementioned adoption - Norm penalty applies specific constraints to the node feature representation, and the specific formula is as follows:

[0025]

[0026] Where β is a trade-off parameter, Z i is the node feature representation of the i-th node, and N is the total number of nodes.

[0027] Furthermore, in the multi-view decoder, a mapping function is added to change the mapping range of node feature representation.

[0028] Furthermore, the mapping function is specifically as follows:

[0029]

[0030] Where x represents the node features.

[0031] A second aspect of the present invention provides a multi-view clustering system based on a graph attention autoencoder.

[0032] A multi-view clustering system based on a graph attention autoencoder includes a view selection module, a feature representation module, a feature constraint module, and a cluster prediction module.

[0033] The view selection module is configured to select the view with the most information from different views of the same group of nodes.

[0034] The feature representation module is configured to: learn the graph structure and node content based on the view and node content information with the most information, and obtain node feature representations by using a graph attention encoder;

[0035] The feature constraint module is configured to: adopt - Norm penalty applies specific constraints to the node feature representation, resulting in a constrained node feature representation;

[0036] The clustering prediction module is configured to input the constrained node feature representations into the self-optimizing clustering module for clustering to obtain the final clustering result.

[0037] Specifically, by adding a multi-view decoder, the constrained node feature representation is reconstructed, and the reconstruction loss is used to train the graph attention encoder.

[0038] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps in the multi-view clustering method based on a graph attention autoencoder as described in the first aspect of the present invention.

[0039] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps in the multi-view clustering method based on a graph attention autoencoder as described in the first aspect of the present invention.

[0040] The above one or more technical solutions have the following beneficial effects:

[0041] This invention proposes a multi-view clustering method based on graph attention autoencoders (MCBGA), which uses graph attention autoencoders to perform attribute multi-view graph clustering tasks. It effectively integrates multi-view graph structure and content information for deep latent representation learning. This model combines node representation learning and clustering into a unified framework, jointly optimizing embedding learning and graph clustering.

[0042] This invention utilizes - Norm penalty addresses the problem of community-specific distribution in node representations, plays an important role in learning node representations, effectively characterizes clustering structures, and thus improves clustering results.

[0043] Experimental results on four benchmark datasets show that the algorithm of this invention outperforms state-of-the-art graph clustering methods.

[0044] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0045] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0046] Figure 1 This is a model structure diagram of the first embodiment.

[0047] Figure 2 This is a flowchart of the method in the first embodiment.

[0048] Figure 3 This is a comparison diagram of the mapping functions in the first embodiment.

[0049] Figure 4 This is a performance comparison chart on the ACM dataset for the first embodiment.

[0050] Figure 5 This is a performance comparison chart on the DBLP dataset for the first embodiment.

[0051] Figure 6 This is a performance comparison chart on the IMDB dataset for the first embodiment.

[0052] Figure 7 Add an MCBGA performance graph to the first embodiment.

[0053] Figure 8 This is a system structure diagram of the second embodiment. Detailed Implementation

[0054] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0055] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the invention; unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0056] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0057] This invention proposes a multi-view clustering method based on a graph attention autoencoder. Using a graph attention autoencoder, it simultaneously reconstructs the graph structure and node content, ensuring that the latent representation effectively preserves the graph structure and content information of the nodes. The core idea of ​​this invention is as follows:

[0058] Using a graph attention autoencoder, shared representations are extracted from the graph view and content data with the most information. Then, all views are reconstructed using the shared representations. The graph attention autoencoder consists of a graph attention encoder and multiple decoders. Specifically, it learns node representations using the multi-view graph structure and node content, learns node representations from the view with the most information through a graph attention encoder, and reconstructs all views through multiple multi-view decoders.

[0059] To eliminate the uncertainty in post-processing clustering operations, a clustering activation function was introduced, which facilitates better node representation learning and node clustering; the inventors discovered... - Norm penalty plays an important role in representing the community-specific distribution of graph structure data in dimensional space. Applying it to the learning of node representations can effectively characterize the clustering structure, thereby improving the clustering results.

[0060] In addition, a self-training clustering objective is designed to make the current clustering distribution tend to be more suitable for the target distribution of the clustering task; by jointly optimizing the reconstruction loss and the clustering loss, the model can simultaneously optimize node embedding and clustering, and improve each other in a unified framework.

[0061] Example 1

[0062] This embodiment discloses a multi-view clustering method based on a graph attention autoencoder;

[0063] like Figure 1 As shown, in order to simultaneously represent a multi-view diagram structure A within a unified framework (1),...,A (M) For node content X, this embodiment proposes a novel graph embedding clustering model, namely a graph attention autoencoder, which includes a graph attention encoder and a multi-view decoder. The graph attention encoder is shared by all views, extracting shared representations from a view graph structure and content data, and a multi-view decoder is designed to reconstruct the multi-view graph data of the shared representations. The reconstruction loss is used to train the graph attention encoder.

[0064] Given a graph G = (V, E1, ..., E M ,X), Represents a set of nodes, e ij (m) ∈E M To represent the relationship between nodes i and j in the m-th view, the topological structure of graph G can be represented by an adjacency matrix set. Let A represent this. (m) The adjacency matrix representing the m-th view, if but otherwise X = {x1,...,x} n} is the attribute value, representing the content information of the node.

[0065] Multi-view clustering methods based on graph attention autoencoders, such as Figure 2 As shown, it includes:

[0066] Step S01: Select the view with the most information from different views of the same group of nodes;

[0067] Different views represent the relationships between the same group of nodes in different aspects, and the content information is shared by all views, with shared information existing between views. Furthermore, in many scenarios, there is usually a view with the most information that controls the performance of the community. Therefore, the content information and the shared information between multiple views are extracted from the view with the most information and the content data, and then used to reconstruct the view.

[0068] We select the view A with the most information and the node content information X as input to reconstruct all views. Here, we use a heuristic metric—modularity—to select the view with the most information based on the result of the modularity function Q. Specifically:

[0069] Each single-view adjacency matrix A (m) The node content information X is input into the graph neural network (GNN) to learn node embeddings;

[0070] The learned embeddings are subjected to k-means to obtain their clustering index;

[0071] Based on clustering index and adjacency matrix A (m)Calculate the modular score for each view and select the view A with the highest score as the view with the most information.

[0072] Modularity is used because it provides an objective metric for evaluating cluster structure.

[0073] Step S02: Based on the view and node content information with the most information, use the trained graph attention encoder to learn the graph structure and node content to obtain node feature representations;

[0074] This embodiment employs a two-layer nonlinear graph attention encoder to map node content information X and view A to obtain potential node feature representation Z. Furthermore, it learns the neighbor representation of each node by paying attention to its neighbors, thus combining the node content information with the graph structure in the potential node feature representation.

[0075] The most direct strategy for learning neighbor representations is to integrate a node representation equally with all its neighbors; however, to measure the importance of various neighbors, the layer-wise graph attention strategy in the graph attention encoder assigns different weights to neighbor representations, representing the importance of neighbor nodes to the current node, specifically:

[0076]

[0077] Among them, z i h N is the node feature representation of node i. i Represents the neighbors of node i; μ ij Let represent the attention coefficient, indicating the importance of neighbor node j to node i, and ω be a non-linear activation function.

[0078] To calculate the attention coefficient μ ij The importance of node j was measured from two aspects: topological structure and attribute values; in terms of attribute values, the attention coefficient μ ij It can be represented as x i x j and weight vector Connected single-layer feedforward neural network:

[0079]

[0080] Attention coefficient μ ij The softmax function is typically used for all neighborhoods j∈N. i Normalize them to make them easy to compare between nodes:

[0081]

[0082] Then, by adding the topology P and the activation function ξ, the attention coefficients are finally expressed as:

[0083]

[0084] The graph attention encoder stacks two graph attention layers, and x i =z i 0 As input:

[0085]

[0086]

[0087] In this way, the encoder encodes both the graph structure and node content into node feature representations, i.e., z. i =z i (2) .

[0088] It should be noted that the last layer of the encoder uses ReLU as the activation function to ensure that the inner product between any two points is non-negative.

[0089] Step S03: Use - Norm penalty applies specific constraints to the node feature representation, resulting in a constrained node feature representation;

[0090] Because the communities represented by nodes have a specific distribution, with different communities distributed across different feature dimensions, the feature distribution of nodes becomes highly chaotic. To ensure that the node feature representation Z can effectively characterize the community structure—in other words, to make the node feature representation Z more discriminative—a certain approach is adopted for Z. Norm penalty, resulting in constrained node feature representations This constraint forces the use of a graph attention encoder to capture the differences in the latent space across different clusters.

[0091] To better utilize the cluster structure, adopt - Norm penalty applies specific constraints to the node feature representation Z in order to achieve better clustering results. - Norm penalty, as a community-specific constraint, is defined as:

[0092]

[0093] Where β is a trade-off parameter, Z i Z is the node feature representation of the i-th node, and N is the total number of nodes; using formula (7), Z i square Different elements in the -norm compete for survival, and Z iAt least one element survives (the remainder is non-zero). By doing so, some discriminative features are preserved for each community, providing some flexibility in learning node representations.

[0094] Step S04: Input the constrained node feature representation into the self-optimizing clustering module for clustering to obtain the final clustering result;

[0095] A self-optimizing clustering module is constructed using K-means. The constrained node feature representation is input into K-means for clustering, and the final clustering result, i.e., the community distribution of the nodes, is output.

[0096] To ensure the supervised graph attention encoder accurately extracts the shared representation of all views, a multi-view decoder is added, which extracts the constrained node feature representations. Reconstructing multi-view data The graph attention encoder is trained using a reconstruction loss. For example... Figure 1 As shown in the multi-view decoder section, the decoder consists of M specific view decoders. Composition, Predictive View A (m) Does a link exist between two nodes, where W... m ∈R D×D View A (m) Specific weights; the multi-view decoder is actually a graph-based embedding multi-view link prediction layer, represented as:

[0097]

[0098] Where σ is the activation function of the multi-view decoder. It is the constrained node feature representation. yes The transpose of .

[0099] Assume all inner products are non-negative, i.e., ZZ T ≥0; because the ReLU activation function of the last coding layer of the graph attention encoder is non-negative, Z≥0, σ(ZZ) T ) will be located In this context, it cannot be considered as a probability.

[0100] To solve this problem, we need to find a mapping function φ that maps [0,+∞) to (-∞,+∞) such that σ(φ(ZZ)) = 0. T The output effective probability, and the ideal mapping function should satisfy the requirement that when the input is large enough, the ideal φ should approximate y = x, such as... Figure 3As shown, if the mapping function is log(x), log(x) is too insensitive to large values, which means that σ(φ(X)) will slowly approach 1 as x increases. Therefore, this embodiment designs the mapping function as follows:

[0101]

[0102] from Figure 3 It can be observed that as x increases, the mapping function shown in formula (9) approaches y = x very quickly, and when x approaches 0, the function rapidly approaches -∞, which meets the conditions of an ideal mapping function.

[0103] Based on the above, the multi-view decoder can be represented as follows:

[0104]

[0105] in, It is the reconstructed view.

[0106] The graph attention autoencoder, which includes both a graph attention encoder and a multi-view decoder, is trained holistically to better learn the graph structure and content information between multiple views. The overall objective function for training is defined as follows: (This is the goal of the training process, which optimizes the clustering performance of the graph attention autoencoder.)

[0107] L = L R +λL C +L norm (11)

[0108] Among them, L R and L C These are reconstruction loss and clustering loss, L norm It is for Z. - Norm penalty serves as a community-specific constraint, with λ≥0 being a coefficient that controls the balance between the two.

[0109] For graph attention autoencoders, the sum of reconstruction errors for each view data is minimized using the following method:

[0110]

[0111] in, For the reconstruction loss of view m, L R For the reconstruction loss of all views, A (m) , These are the original view and the reconstructed view, respectively. Since the decoder adopts a multi-view structure, the gradients of the multiple decoders will be propagated through the information graph encoder during backpropagation. Therefore, during forward propagation, the graph encoder will extract the shared representation of all views.

[0112] In addition to optimizing the reconstruction error, the clustering loss is also controlled. To this end, a self-optimizing clustering module is added. The constrained node feature representation is input into the self-optimizing clustering module to minimize the following objective:

[0113]

[0114] Where KL(·|·) is the Kullback-Leibler divergence between the two distributions, and Q is the distribution of soft labels to indicate the constrained node feature representation of node i. and cluster center u j Similarities between them:

[0115]

[0116] q ij It can be viewed as a soft clustering distribution of each node, where p in formula (13) ij The target distribution is defined as:

[0117]

[0118] Soft assignments with high probability (nodes close to the community center) are considered trustworthy in Q, so the target distribution P raises Q to the quadratic to emphasize the role of those "confident assignments". Then the clustering loss forces the current distribution Q to approximate the target distribution P, thus setting these "confident assignments" as soft labels to supervise the embedding learning of Q.

[0119] The target distribution P acts as the "true label" during training, but it also depends on the current soft assignment Q updated at each iteration. Updating P with Q at each iteration is dangerous because the constant changes in the target can hinder learning and convergence. In order to avoid instability in the self-optimization process, this embodiment updates P every 5 iterations in the experiment.

[0120] Minimizing the clustering loss helps the autoencoder manipulate the embedding space by leveraging the features of the embeddings themselves and the scattered embedding points, thereby achieving better clustering performance.

[0121] Using the target distribution P and soft label distribution Q mentioned above, update the model weights and cluster centers as follows:

[0122] (1) With a fixed target distribution P, given N samples, L C Relative to cluster center μ j The gradient can be calculated as:

[0123]

[0124] (2) Given the learning rate λ, update μ j for:

[0125]

[0126] (3) Update the weights of the specific decoder for the m-th view:

[0127]

[0128] As can be seen, W m The update is only related to the reconstruction loss of view m. Therefore, the weights of the view-specific decoder can capture view-specific local structural information, and thus the weights of the encoder can extract the shared representation of all views.

[0129] The effectiveness of the method in this embodiment was verified through comparative experiments.

[0130] Experimental data

[0131] ACM: This is a paper network in the ACM dataset, constructed using a two-view graph that leverages co-paper (two papers written by the same author) and co-subject (two papers covering the same topic) relationships. Paper features are elements of a bag-of-words represented by keywords.

[0132] DBLP consists of three graphs: a collaboration graph, a paper citation graph, and a paper co-citation graph. The collaboration graph has 2401 author nodes and 8703 edges; the paper citation graph has 6000 paper nodes and 10003 edges; the paper co-citation graph has 6000 paper nodes and 141996 edges (two nodes are connected if they cite a common paper); authors and papers are linked by 32048 author identities; links between papers in the citation and co-citation graphs are based on paper identities; all authors and papers are involved in three clusters representing research areas: artificial intelligence, computer graphics, and computer networks.

[0133] IMDB: This is a movie network from the IMDB dataset. It uses co-actors (movies starring the same actor) and co-directors (movies directed by the same director) relationships to build a multi-view. Movie features correspond to a set of words representing the plot.

[0134] The clustering performance of the method in this embodiment was evaluated on three datasets: ACM, DBLP, and IMDB. Brief statistics for the three datasets are shown in Table 1.

[0135] Table 1 Dataset Information

[0136]

[0137]

[0138] • Parameter settings and evaluation indicators

[0139] Four commonly used metrics are used: Accuracy (ACC), F-score (F1), Normalized Mutual Information (NMI), and Alternating Land Index (ARI). The ARI value ranges from -1 to 1, with a larger value being better, reflecting the degree of overlap between the two partitions. For each metric, a larger value means better clustering results.

[0140] For the ACM dataset, which is relatively small, it was trained for 250 iterations. For the DBLP and IMDB datasets, all autoencoder models were trained for 1000 iterations and optimized using the Adam algorithm. The learning rate λ of the autoencoder was set to 0.001, and the dimension of all embedding methods was set to 32. The convergence threshold for MCBGA was set to 0.1%, and the update period T = 20. For the remaining methods, the settings described in the respective papers were retained. Since all clustering algorithms depend on initialization, all methods were repeated 10 times using random initialization.

[0141] • Comparison Method

[0142] GAE: A single-view graph autoencoder method;

[0143] X-avg: In order to take advantage of multiple views of the network, the X method is used to learn the node representations on each view, and then all learned representations are averaged.

[0144] MNE: A scalable multi-view network embedding model that uses the multi-view graph adjacency matrix as input for all multi-view graph embedding / clustering methods;

[0145] RMSC: A robust multi-view spectral clustering method based on low-rank and sparse decomposition;

[0146] PwMC is a parametric weighted multi-view graph clustering method.

[0147] SwMC: A self-weighted multi-view graph clustering method;

[0148] O2MA: A variant of O2MAC that does not include clustering loss in the objective function;

[0149] O2MAC: Proposes an attribute-based multi-view graph clustering method;

[0150] MCBGA: This embodiment proposes a multi-view clustering method based on graph attention autoencoders.

[0151] Experimental Results

[0152] To evaluate the efficiency of the proposed method, the clustering performance on the ACM, DBLP, and IMDB datasets was analyzed. Tables 2, 3, and 4 summarize the experimental results on the three benchmark datasets, where bold values ​​indicate the best performance. As shown in Tables 2, 3, and 4, the experimental results demonstrate that the method in this embodiment significantly outperforms all other methods in most evaluation metrics.

[0153] Table 2 Experimental results on the ACM dataset

[0154] method ACC F1 NMI ARI GAE 0.8216 0.8225 0.4914 0.5444 GAE-avg 0.6990 0.7025 0.4771 0.4378 MNE 0.6370 0.6479 0.2999 0.2486 RMSC 0.6315 0.5746 0.3973 0.3312 PwMC 0.4162 0.3783 0.0332 0.0395 SwMC 0.3831 0.4709 0.0838 0.0187 O2MA 0.8880 0.8894 0.6515 0.6987 O2MAC 0.9042* 0.9053* 0.6923* 0.7394* MCBGA 0.9102 0.9223 0.7052 0.7451

[0155] Table 3 Experimental results on the DBLP dataset

[0156] method ACC F1 NMI ARI GAE 0.8859 0.8743 0.6925 0.7410 GAE-avg 0.5558 0.5418 0.3072 0.2577 MNE - - - - RMSC 0.8994 0.8248 0.7111 0.7647 PwMC 0.3253 0.2808 0.0190 0.0159 SwMC 0.6538 0.5602 0.3760 0.3800 O2MA 0.9040 0.8976 0.7257 0.7705 O2MAC 0.9074* 0.9013* 0.7287* 0.7780* MCBGA 0.9156 0.9047 0.7365 0.7789

[0157] Table 4 Experimental results of the IMDB dataset

[0158]

[0159]

[0160] In this diagram, "*" indicates the best performance of the baseline, the best results of all methods are shown in bold, and "-" indicates that the methods on this dataset are out of memory.

[0161] The experimental results are visualized as follows: Figure 4-5 As shown in the visualization, the MCBGA results outperform almost all baseline methods, indicating that the model proposed in this embodiment is effective. Then, by comparing the results of MCBGA with GAE-avg, O2MA, and O2MAC, it is concluded that the MCBGA of this embodiment is a more effective graph neural network for fusing multi-view information. Furthermore, compared to O2MAC, the MCBGA of this embodiment achieves better results on three datasets. For example, on the ACM dataset, the MCBGA of this embodiment improves the four metrics of O2MAC (ACC, F1, NMI, and ARI) by 0.6%, 1.8%, 1.1%, and 0.7%, respectively. This may be because the MCBGA of this embodiment utilizes… - Norm penalty addresses the community-specific distribution of node representations and plays a crucial role in learning node representations. It can effectively characterize clustering structures, thereby improving clustering results. Compared to O2MA, the better results of MCBGA in this embodiment on three datasets demonstrate that self-trained clustering objectives can further improve clustering performance after effective pre-training. It is also noted that the results of all baselines on the IMDB dataset are lower than those on the ACM and DBLP datasets because it is difficult to obtain "highly confident" nodes on IMDB. In this case, these "highly confident" nodes may confine low-confidence nodes to incorrect clusters.

[0162] The MCBGA algorithm proposed in this embodiment achieves good performance on both the ACM and DBLP datasets, demonstrating its superiority in node clustering tasks. For example, on the ACM dataset, the clustering performance of the MCBGA algorithm in this embodiment is significantly improved compared to the GAE method. This is because the MCBGA algorithm in this embodiment utilizes complementary information embedded in multi-view data, which single-view methods do not. In general, single-view-based clustering methods are inferior to multi-view-based clustering methods because multi-view methods can utilize complementary information embedded in multi-view data, while single-view methods cannot.

[0163] Based on the experimental results, the following conclusions can be drawn: First, the embedding method is significantly superior to other methods; therefore, graph embedding is a promising approach for solving graph clustering problems. Second, the deep learning method (GAE) achieved more competitive results than other baselines; however, it can only utilize the graphical and content information of a single view. A well-designed deep neural network that integrates multiple graph views has the potential to achieve good results.

[0164] Ablation experiment

[0165] Ablation studies were conducted to further analyze the importance of each module in the framework proposed in this embodiment.

[0166] Table 5 Ablation experiments on the three datasets

[0167]

[0168] Setting whether or not - Norm penalty was used for comparative experiments. As can be seen from Table 5, when setting The experimental results of the norm penalty are better, indicating that the proposed method in this embodiment is superior. - Norm penalty helps learn better discriminative representations of potential nodes for node clustering tasks.

[0169] Secondly, to verify the necessity of using modularity to select infographic views, each infographic view was used as input to MCBGA to reconstruct all infographic views on the three datasets, and their results on the graph clustering task are shown in Table 6. It can be observed that if infographic views with higher modularity values ​​are input into the encoder, the model will obtain better results, verifying that modularity is a feasible solution for infographic view selection.

[0170] Table 6 shows the aggregated results across different input views.

[0171]

[0172]

[0173] The selected views used in the experiment are shown in bold.

[0174] In the model proposed in this embodiment, multiple graph views are fused to improve clustering performance. To further investigate the impact of multiple views on learning embeddings for clustering tasks, the performance of MCBGA is carefully studied by adding graph views one by one to the DBLP dataset. These three views are co-conference, co-term, and co-paper, which are added to the model in sequence. Figure 7 The results of these four metrics demonstrate the performance of the MCBGA with additional views. The results show that the performance of the model proposed in this embodiment steadily improves as views are added one by one. Therefore, MCBGA provides a flexible framework to utilize more graph views.

[0175] Example 2

[0176] This embodiment discloses a multi-view clustering system based on a graph attention autoencoder;

[0177] like Figure 8 As shown, the multi-view clustering system based on graph attention autoencoder includes a view selection module, a feature representation module, a feature constraint module, and a clustering prediction module:

[0178] The view selection module is configured to select the view with the most information from different views of the same group of nodes.

[0179] The feature representation module is configured to: learn the graph structure and node content based on the view and node content information with the most information, and obtain node feature representations by using a graph attention encoder;

[0180] The feature constraint module is configured to: adopt - Norm penalty applies specific constraints to the node feature representation, resulting in a constrained node feature representation;

[0181] The clustering prediction module is configured to input the constrained node feature representations into the multi-view decoder for prediction, and obtain the final clustering result.

[0182] Example 3

[0183] The purpose of this embodiment is to provide a computer-readable storage medium.

[0184] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps in the multi-view clustering method based on a graph attention autoencoder as described in Embodiment 1 of this disclosure.

[0185] Example 4

[0186] The purpose of this embodiment is to provide an electronic device.

[0187] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the multi-view clustering method based on a graph attention autoencoder as described in Embodiment 1 of this disclosure.

[0188] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-view clustering method based on graph attention autoencoders, characterized in that, include: Select the view with the most information from different views of the same set of nodes; The different views include a collaboration graph, a paper citation graph, and a paper co-citation graph; The collaboration graph consists of author nodes and edges, the paper citation graph consists of paper nodes and edges, and the paper co-citation graph consists of paper nodes and edges. Authors and papers are linked by author identity, and the links between papers in the citation graph and co-citation graph are based on the paper identity. Based on the view and node content information with the most information, a trained graph attention encoder is used to learn the graph structure and node content to obtain node feature representations. A two-layer nonlinear graph attention encoder is used to map node content information X and view A to obtain potential node feature representation Z. Furthermore, by paying attention to its neighbors, the neighbor representation of each node is learned, combining the node content information with the graph structure in the potential node feature representation. In the layer-by-layer graph attention strategy of the graph attention encoder, different weights are assigned to the neighbor representations, representing the importance of neighbor nodes to the current node. Specifically: ; in, It is the node feature representation of node i. Indicate the neighbors of node i; The attention coefficient represents the importance of neighbor node j to node i. It is a non-linear activation function; To calculate the attention coefficient The importance of node j was measured from two aspects: topological structure and attribute values; in terms of attribute values, the attention coefficient... Represented as , and weight vector Connected single-layer feedforward neural network: ; Attention coefficient Use the softmax function to evaluate all neighborhoods. Normalize the above to make all neighborhoods Comparison between nodes: ; Then add the topology P and the activation function. The attention coefficient is then ultimately expressed as: ； The graph attention encoder stacks two graph attention layers, As input: ; ; In this way, the encoder encodes both the graph structure and node content into node feature representations, i.e. ; use The 1,2-norm penalty applies a specific constraint to the node feature representation, resulting in a constrained node feature representation. The constrained node feature representations are input into the self-optimizing clustering module for clustering, resulting in the final clustering results, which involve three clusters. Specifically, by adding a multi-view decoder, the constrained node feature representation is reconstructed, and the reconstruction loss is used to train the graph attention encoder.

2. The multi-view clustering method based on graph attention autoencoder as described in claim 1, characterized in that, The view with the most information is selected based on the clustering index and adjacency matrix, by calculating the modular score of each graph view and selecting the graph view with the highest score as the view with the most information.

3. The multi-view clustering method based on graph attention autoencoder as described in claim 1, characterized in that, The graph attention encoder also includes learning the importance of neighboring nodes.

4. The multi-view clustering method based on graph attention autoencoder as described in claim 3, characterized in that, The learning of the importance of neighboring nodes is achieved by assigning different weights to neighbors in the layer-by-layer graph attention strategy of the graph attention encoder, representing the importance of neighboring nodes to the current node.

5. The multi-view clustering method based on graph attention autoencoder as described in claim 1, characterized in that, The use The 1,2-norm penalty applies specific constraints to the node feature representation, and the specific formula is as follows: Where β is a trade-off parameter, Z i is the node feature representation of the i-th node, and N is the total number of nodes.

6. The multi-view clustering method based on graph attention autoencoder as described in claim 1, characterized in that, In the multi-view decoder, a mapping function is added to change the mapping range of node feature representation.

7. The multi-view clustering method based on graph attention autoencoder as described in claim 6, characterized in that, The mapping function is specifically: Where x represents the node features.

8. A multi-view clustering system based on a graph attention autoencoder, characterized in that, The multi-view clustering method based on graph attention autoencoder as described in any one of claims 1-7 includes a view selection module, a feature representation module, a feature constraint module, and a clustering prediction module. The view selection module is configured to select the view with the most information from different views of the same group of nodes. The feature representation module is configured to: learn the graph structure and node content based on the view and node content information with the most information, and obtain node feature representations by using a graph attention encoder; The feature constraint module is configured to: adopt The 1,2-norm penalty applies a specific constraint to the node feature representation, resulting in a constrained node feature representation. The clustering prediction module is configured to input the constrained node feature representations into the self-optimizing clustering module for clustering to obtain the final clustering result. Among them, a multi-view decoder is added to reconstruct the constrained node feature representation, and the reconstruction loss is used to train the graph attention encoder.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the multi-view clustering method based on graph attention autoencoder as described in any one of claims 1-7.

10. An electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the multi-view clustering method based on graph attention autoencoder as described in any one of claims 1-7.

Images