Depth attribute graph clustering method

Abstract

The invention relates to the field of depth attribute graphs, and particularly discloses a depth attribute graph clustering method, which adopts a dual-drive network framework DynStuct-Cluster, fuses a dual-branch structure of an auto-encoder AE and a graph auto-encoder GAE, and clusters an attribute graph through joint optimization of a graph structure and a clustering process. The dual-drive network framework DynStuct-Cluster is composed of an auto-encoder AE, a graph convolutional network GCN, a dynamic graph structure optimization DGR, a dynamic prototype comparative learning DPCL and a self-supervised optimization module. According to the method, the discrimination of clustering characterization is further improved, and adaptive optimization of the graph structure is effectively promoted.

Description

Technical Field

[0001] The present invention relates to the technical field of deep attribute graphs, and in particular to a deep attribute graph clustering method. Background Art

[0002] Attributed graph clustering, a core task in graph data analysis, aims to achieve semantic segmentation by fusing node features with non-Euclidean structural information. It is widely used in fields such as object detection and biological network analysis. However, the heterogeneity of node attributes and the complexity of graph structures make it difficult for existing methods to simultaneously capture both structural and semantic information. Current deep learning-based clustering models (such as SDCN, DFCN, and CAEGCN) have made progress in feature extraction and structural modeling by fusing autoencoders (AEs) with graph convolutional networks (GCNs), but they still face two major challenges:

[0003] Insufficient adaptability to dynamic structures: Traditional methods rely on static adjacency matrices, making it difficult to model the dynamic evolution of semantic relationships. While some work has attempted to construct dynamic similarity graphs through node embedding dot products, the lack of original topological constraints can easily lead to semantic drift (for example, GRACE introduced a graph enhancement strategy but did not constrain the direction of topological evolution). GCN's over-smoothing problem further limits its ability to model complex structures. Although DFCN, DAGC, etc. have alleviated this problem through attention mechanisms or dual self-supervision strategies, they have not addressed the fundamental need for semantic-structural co-evolution.

[0004] Lack of semantic alignment mechanisms: Existing methods lack direct alignment between embedded representations and semantic prototypes, leading to prototype drift. While contrastive learning methods (such as GCA and SCAGC) improve discriminability through instance-level comparison, they do not explicitly model global cluster structure, resulting in an imbalance between intra-class closeness and inter-class separability. While methods like DCRN attempt to reduce representational redundancy, they still fail to establish a direct mapping between the embedding space and semantic prototypes.

[0005] Recent research (such as HSAN and CCGC) has improved comparison quality through twin encoders and high-confidence sample mining, but global cluster structure modeling still relies on heuristic design. Existing solutions focus primarily on local structure optimization and lack joint constraints on the dynamic changes in semantic distribution and global cluster structure, making it difficult to maintain structural-semantic consistency in complex scenarios. Future research requires exploring a joint optimization framework for dynamic graph modeling and prototype alignment to achieve structural adaptability and semantic interpretability in attribute graph clustering. Summary of the Invention

[0006] In response to the above-mentioned problems in the prior art, the present invention proposes a deep attribute graph clustering method, which can improve the discriminability of cluster representation and effectively promote the adaptive optimization of graph structure.

[0007] To achieve the above objectives, the present invention proposes a deep attribute graph clustering method, a deep attribute graph clustering method, which adopts a dual-drive network framework DynStruct-Cluster, integrates the dual-branch structure of autoencoder AE and graph autoencoder GAE, and clusters attribute graphs by jointly optimizing graph structure and clustering process; the dual-drive network framework DynStruct-Cluster consists of autoencoder AE, graph convolutional network GCN, dynamic graph structure optimization DGR, dynamic prototype contrast learning DPCL and self-supervised optimization module.

[0008] Preferably, the steps of using the dual-driver network framework DynStruct-Cluster to perform deep attribute graph clustering are:

[0009] The S1 and DGR modules dynamically optimize the structure of the original adjacency matrix and update the adjacency matrix according to the current node features and the structure of the graph in each training iteration. The input data consists of the node feature matrix X and the adjacency matrix A.

[0010] S2, AE and GCN modules learn two heterogeneous representations from the attributes and structure of the graph respectively. Each layer of features of the AE encoder is fused with each layer of features of the GCN to obtain the final fused features;

[0011] S3, perform prototype comparison through the DPCL module and calculate the contrast loss;

[0012] S4. Generate soft assignments Q and Q′ of clusters through the latent representation and the final fused features, where the target distribution p is generated by Q and guides the self-supervised clustering process.

[0013] Preferably, the autoencoder AE module learns low-dimensional representations from node attributes and generates a potential representation Z to assist downstream node clustering tasks; the autoencoder AE adopts a symmetrical structure and consists of an encoder and a decoder. The encoder maps the original data to a low-dimensional latent space by stacking multiple layers of nonlinear transformations, and the decoder reconstructs the original feature data to retain the main structural information of the input data.

[0014] Preferably, the specific process of encoding and decoding the autoencoder AE and performing optimization training to generate the potential representation Z is:

[0015] Step 1: In the encoding stage, the original input data is Where N is the number of samples, d is the input feature dimension, and assuming that AE consists of L layers, the hidden representation of the lth layer is:

[0016]

[0017] Where W e(l) and denote the weight matrix and bias vector of the corresponding layer respectively, σ(·) is the activation function ReLU of the fully connected layer;

[0018] Step 2: In the decoding stage, the decoder reconstructs the input layer by layer, and the decoder output formula is:

[0019]

[0020] The final output is the reconstruction result

[0021] Step 3: The model is trained by minimizing the difference between the input and the reconstruction. The optimization goal is:

[0022]

[0023] Where x i and Represent the original features and reconstructed features of the i-th sample, respectively, ‖·‖ F Represents the Frobenius norm, and finally the potential representation Z obtained by the last layer output of the encoder, where Z = Y (L) Serves as input for subsequent graph structure learning and clustering optimization.

[0024] Preferably, the graph convolutional network (GCN) module captures the topological relationship between nodes and their neighbors and fuses the topological information of GCN with the feature representation of AE.

[0025] Preferably, the fusion process of the topological information of GCN and the feature representation of AE is as follows: a linear weighted fusion strategy is adopted to unify the structural information and attribute information into a unified model, and the calculation expression is:

[0026]

[0027] Where, is the output of the AE layer l, α = 0.5;

[0028] Combine AE and GCN layer by layer to obtain a new fusion feature representation. The calculation expression is:

[0029]

[0030] Where, represents the adjacency matrix after adding the self-loop, Its degree matrix, W (l) is the weight matrix of the lth layer, σ is the nonlinear activation function;

[0031] The last layer of GCN output is passed through Softmax to generate a cluster probability distribution, and a multi-category probability distribution matrix is ​​obtained, in which each row is a probability distribution of a sample. The calculation formula is:

[0032] Q′=softmax(H (L) );

[0033] Where Q′ represents each element Q′ ij Represents the probability that sample i belongs to category j.

[0034] Preferably, the dynamic graph structure optimization module DGR dynamically optimizes the graph structure through three stages: dynamic similarity modeling, adaptive sparsification, and graph structure fusion. The dynamic similarity modeling stage calculates the dynamic similarity between nodes based on node attributes and structure information. The calculation formula is:

[0035]

[0036] Where, X i represents the input feature of the i-th node, X is the original feature of the current node, and S is the dynamic similarity graph;

[0037] The adaptive sparsification stage performs a row-by-row Top-k screening on the similarity matrix, retaining the most relevant K neighbors of each node while excluding self-loops, generating a sparse similarity graph, and symmetric normalizing the sparse similarity graph to construct an optimized adjacency matrix. The calculation formula is:

[0038]

[0039] Where, is a sparse similarity graph, A s is the adjacency matrix, D is the degree matrix, and its diagonal elements are defined as

[0040] The graph structure fusion stage fuses the optimized graph structure with the original graph structure to obtain the final graph structure and participate in the feature propagation process of GCN. The optimized propagation form is expressed as:

[0041]

[0042] in, is the final adjacency matrix, α∈[0,1] is the learnable fusion coefficient.

[0043] Preferably, the dynamic prototype contrastive learning module DPCL utilizes a prototype clustering mechanism to enhance the discrimination of sample representation, optimizes the prototype representation through a momentum update strategy, and enhances intra-class consistency and inter-class discrimination; the prototype clustering mechanism uses the central representation of each category as the prototype, and performs cluster assignment by calculating the distance between the sample and the prototype; the momentum update strategy dynamically updates the prototype representation based on the historical prototype representation and the current sample representation.

[0044] Preferably, the specific process of forming a cluster structure using the dynamic prototype contrastive learning module DPCL is:

[0045] First, the K-means algorithm is used to cluster the high-order feature representation based on the GCN output to generate preliminary pseudo labels. Based on these pseudo labels, the prototype vector of each class is calculated:

[0046]

[0047] Where C k is the prototype vector of each class, Represents the set of nodes currently assigned pseudo labels as the Kth class, V k is the number of nodes in the set, h i Represents the feature representation of node i, which is the final fusion representation of the GCN module;

[0048] Secondly, the momentum update strategy is used to dynamically adjust the prototype vector of each class. The update formula is as follows:

[0049]

[0050] Where β∈[0,1] is a fixed momentum coefficient;

[0051] Finally, maximize the similarity of each sample with its prototype and minimize its similarity with prototypes of other categories, given the temperature parameter τ , calculate the similarity matrix between all nodes and the prototype, and define the prototype contrast loss of each sample as the normalized cross entropy to form a more discriminative clustering structure. The calculation formula is:

[0052]

[0053] Preferably, the self-supervisory optimization module optimizes model parameters through self-supervisory signals; the self-supervisory signals include but are not limited to feature reconstruction loss, cluster optimization loss, cluster distribution consistency constraint and dynamic prototype comparison loss; the feature reconstruction loss, cluster optimization loss, cluster distribution consistency constraint and dynamic prototype comparison loss constitute a total loss function, which is expressed as:

[0054] L=L res+ λ 1L clu + λ 2L cess + λ 3L proto ;

[0055] Where, L res is the reconstruction loss of the feature, L clu Optimize the loss for clustering, L cess is the cluster distribution consistency constraint, L DPCL is the dynamic prototype contrast loss, and λ1, λ2, and λ3 are hyperparameters.

[0056] Therefore, the present invention proposes a deep attribute graph clustering method, which has the following beneficial effects:

[0057] DynStruct-Cluster effectively alleviates the problems of static structural dependency and semantic alignment through dynamic graph optimization and prototype comparative learning, enhances intra-class compactness and inter-class differentiation, and significantly improves clustering performance on multiple datasets, verifying its effectiveness and robustness.

[0058] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 The cosine similarity moments of the latent space between SDCN and the present invention are visualized on the ACM and DBLP datasets;

[0060] Figure 2 This is a schematic diagram of the DynStruct-Cluster architecture of a deep attribute graph clustering method of the present invention;

[0061] Figure 3 This is a comparison of the clustering performance of DynStruct-Cluster, SDCN, and DFCN on five standard datasets under the same settings of a deep attribute graph clustering method of the present invention;

[0062] Figure 4 is a different k of the deep attribute graph clustering method of the present invention top Compared with the clustering performance of ACM, CiteSeer, and DBLP under fixed λ;

[0063] Figure 5 It is a dynamic optimization process of ACM, CiteSeer, and DBLP that is adaptive under optimal conditions for a deep attribute graph clustering method of the present invention;

[0064] Figure 6 It is the clustering result of a deep attribute graph clustering method with different hyperparameters of the present invention;

[0065] Figure 7 This is an ablation study of five benchmarks for four indicators of a deep attribute graph clustering method of the present invention;

[0066] Figure 8 This paper presents a deep attribute graph clustering method that uses t-SNE to perform two-dimensional visualization of the cluster divisions of AE, SDCN, DFCN, and DynStruct-Cluster on four datasets;

[0067] Figure 9 This is the convergence of DynStruct-Cluster, a deep attribute graph clustering method of the present invention, on five data sets. DETAILED DESCRIPTION

[0068] To make the technical solutions, advantages, and purposes of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are part of the embodiments of the present invention, not all of them. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0069] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0070] like Figure 1 As shown in the figure, the latent space cosine similarity matrix of SDCN and a deep attribute graph clustering method provided by the present invention is visualized on the ACM and DBLP datasets. The red diagonal line in the figure represents the high similarity between nodes of the same type, while the low similarity in the non-diagonal area indicates good separation between different categories. The results show that SDCN has low similarity between nodes of the same type, resulting in intra-class dispersion; while the deep attribute graph clustering method provided by the present invention constructs a tighter similarity structure and forms clear separation boundaries between different categories, significantly improving the clustering discrimination ability.

[0071] like Figure 2 As shown, the present invention provides a deep attribute graph clustering method, which adopts the dual-drive network framework DynStruct-Cluster, integrates the dual-branch structure of autoencoder AE and graph autoencoder GAE, models graph data from two levels: node attributes and graph structure, and clusters attribute graphs by jointly optimizing the graph structure and clustering process.

[0072] The dual-driver network framework DynStruct-Cluster consists of an autoencoder (AE), a graph convolutional network (GCN), dynamic graph structure optimization (DGR), dynamic prototype contrastive learning (DPCL), and a self-supervised optimization module. The steps for deep attribute graph clustering are as follows:

[0073] The S1 and DGR modules dynamically optimize the structure of the original adjacency matrix and update the adjacency matrix according to the current node features and the structure of the graph in each training iteration. The input data consists of the node feature matrix X and the adjacency matrix A.

[0074] S2, AE and GCN modules learn two heterogeneous representations from the attributes and structure of the graph respectively. Each layer of features of the AE encoder is fused with each layer of features of the GCN to obtain the final fused features;

[0075] S3, perform prototype comparison through the DPCL module and calculate the contrast loss;

[0076] S4. Generate soft assignments Q and Q′ of clusters through the latent representation and the final fused features, where the target distribution p is generated by Q and guides the self-supervised clustering process.

[0077] The autoencoder AE module learns low-dimensional representations from node attributes and generates a potential representation Z to assist downstream node clustering tasks; the autoencoder AE adopts a symmetrical structure and consists of two parts: an encoder and a decoder. The encoder maps the original data to a low-dimensional latent space by stacking multiple layers of nonlinear transformations, and the decoder reconstructs the original feature data to retain the main structural information of the input data.

[0078] The specific process of encoding and decoding by the autoencoder AE and performing optimization training to generate the potential representation Z is as follows:

[0079] Step 1: In the encoding stage, the original input data is Where N is the number of samples, d is the input feature dimension, and assuming that AE consists of L layers, the hidden representation of the lth layer is:

[0080]

[0081] Where W e (l) and denote the weight matrix and bias vector of the corresponding layer respectively, σ(·) is the activation function ReLU of the fully connected layer;

[0082] Step 2: In the decoding stage, the decoder reconstructs the input layer by layer, and the decoder output formula is:

[0083]

[0084] The final output is the reconstruction result

[0085] Step 3: The model is trained by minimizing the difference between the input and the reconstruction. The optimization goal is:

[0086]

[0087] Where x i and Represent the original features and reconstructed features of the i-th sample, respectively, ‖·‖ F represents the Frobenius norm, and the final potential representation Z is used as the input for subsequent graph structure learning and clustering optimization.

[0088] The graph convolutional network (GCN) module captures the topological relationship between nodes and their neighbors and fuses the GCN topological information with the feature representation of the automatic feature extraction (AE). The specific process of fusing the GCN topological information with the feature representation of the automatic feature extraction (AE) is as follows:

[0089] A linear weighted fusion strategy is used to unify the structural information and attribute information into a unified model. The calculation expression is:

[0090]

[0091] Where H (l) is the GCN representation of the lth layer, where the first layer H (0) Using the original feature X, is the output of the AE layer l, α = 0.5;

[0092] Combine AE and GCN layer by layer to obtain a new fusion feature representation. The calculation expression is:

[0093]

[0094] Where, represents the adjacency matrix after adding the self-loop, Its degree matrix, W (l) is the weight matrix of the lth layer, σ is the nonlinear activation function;

[0095] The last layer of GCN output is passed through Softmax to generate a cluster probability distribution, and a multi-category probability distribution matrix is ​​obtained, in which each row is a probability distribution of a sample. The calculation formula is:

[0096] Q′=softmax(H (L) );

[0097] Where Q′ represents each element Q′ ij Represents the probability that sample i belongs to category j.

[0098] The dynamic graph structure optimization module DGR dynamically optimizes the graph structure through three stages: dynamic similarity modeling, adaptive sparsification, and graph structure fusion. In the dynamic similarity modeling stage, the dynamic similarity between nodes is calculated based on node attributes and structure information. The calculation formula is:

[0099]

[0100] Where, X i represents the input feature of the i-th node, X is the original feature of the current node, and S is the dynamic similarity graph;

[0101] In the adaptive sparsification stage, the similarity matrix is ​​filtered row by row using the Top-k filter, retaining the most relevant K neighbors of each node while excluding self-loops, generating a sparse similarity graph, and symmetrically normalizing the sparse similarity graph to construct the optimized adjacency matrix. The calculation formula is:

[0102]

[0103]

[0104] Where, is a sparse similarity graph, A s is the adjacency matrix, D is the degree matrix, and its diagonal elements are defined as

[0105] In the graph structure fusion stage, the optimized graph structure is fused with the original graph structure to obtain the final graph structure and participate in the feature propagation process of GCN. The optimized propagation form is expressed as:

[0106]

[0107] in, is the final adjacency matrix, α∈[0,1] is the learnable fusion coefficient.

[0108] The dynamic prototype contrastive learning module (DPCL) uses the prototype clustering mechanism to enhance the discriminability of sample representation, optimizes the prototype representation through the momentum update strategy, and enhances intra-class consistency and inter-class discrimination. The prototype clustering mechanism uses the central representation of each category as the prototype and performs cluster assignment by calculating the distance between the sample and the prototype. The momentum update strategy dynamically updates the prototype representation based on the historical prototype representation and the current sample representation.

[0109] The specific process of forming a clustering structure using the dynamic prototype contrastive learning module DPCL is as follows:

[0110] First, the K-means algorithm is used to cluster the high-order feature representation based on the GCN output to generate preliminary pseudo labels. Based on these pseudo labels, the prototype vector of each class is calculated:

[0111]

[0112] Where C k is the prototype vector of each class, Represents the set of nodes currently assigned pseudo labels as the Kth class, V k is the number of nodes in the set, h i Represents the feature representation of node i, which is the final fusion representation of the GCN module;

[0113] Secondly, the momentum update strategy is used to dynamically adjust the prototype vector of each class. The update formula is as follows:

[0114]

[0115] Where β∈[0,1] is a fixed momentum coefficient;

[0116] Finally, maximize the similarity of each sample with its prototype and minimize its similarity with prototypes of other categories, given the temperature parameter τ , calculate the similarity matrix between all nodes and the prototype, and define the prototype contrast loss of each sample as the normalized cross entropy to form a more discriminative clustering structure. The calculation formula is:

[0117]

[0118] The self-supervised optimization module optimizes model parameters through self-supervisory signals and integrates a self-supervised optimization mechanism. This mechanism combines the autoencoder (AE) and the graph convolutional network (GCN) to promote the collaborative learning of structural and attribute information. The cluster distribution generated by the AE branch is treated as a pseudo-label to guide the GCN branch in learning structure-aware feature representations. Specifically:

[0119] Use AE branch to obtain node representation Z i , and calculate its relationship with the cluster center μ j The similarity between them is used to generate a probability distribution that conforms to the Student's t distribution based on the similarity:

[0120]

[0121] Where q ij represents sample z i Assigned to the cluster center μ j The probability of the current model is regarded as the soft prediction distribution of the sample distribution. The degree of freedom v of the Student'st distribution is set to 1. The entire distribution Q=[q ij ] is regarded as the initial pseudo-label source of the clustering structure;

[0122] Constructing high-confidence cluster target distribution It is used as a pseudo-label to perform self-supervisory optimization on the model, improve the confidence of the pseudo-label and enhance the guidance ability of the clustering target. The calculation process of the target distribution P is as follows:

[0123]

[0124] By squaring the entire distribution Q and normalizing it by column, we can obtain the following objective function by minimizing the KL divergence between and :

[0125]

[0126] Among them, the pseudo label P is obtained by the prediction Q generated by the model itself, and is used in reverse to supervise the update of the model;

[0127] The AE and GCN branches learn node representations from attribute information and structural information respectively, enhance the alignment of the two branches in the clustering structure, use the target distribution P constructed by the AE branch as a guide, supervise the predicted distribution Q′ of the GCN branch, and introduce a consistency constraint loss function:

[0128]

[0129] Self-supervisory signals include but are not limited to feature reconstruction loss, cluster optimization loss, cluster distribution consistency constraint, and dynamic prototype contrast loss.

[0130] The total loss function consists of feature reconstruction loss, cluster optimization loss, cluster distribution consistency constraint and dynamic prototype comparison loss, which is expressed as:

[0131] L=L res + λ 1L clu + λ 2L cess + λ 3L proto ;

[0132] Where, L res is the reconstruction loss of the feature, L clu Optimize the loss for clustering, L cess is the cluster distribution consistency constraint, L DPCL is the dynamic prototype contrast loss, and λ1, λ2, and λ3 are hyperparameters. By optimizing this objective function, the overall network architecture can be trained in a self-supervised manner, thereby achieving optimization of the clustering task.

[0133] Based on the complexity analysis of a deep attribute graph clustering method of the present invention, the time complexity is mainly composed of autoencoder (AE), graph convolutional network (GCN), dynamic graph structure learning, clustering and self-supervision modules.

[0134] The complexity analysis is as follows: let N be the number of nodes, |E| be the number of edges in the graph, d0 be the input feature dimension, d L is the dimension of the potential representation, d z The dimension in the latent space, L is the number of layers of the encoder and GCN, K is the number of clusters, k top is the number of Top-k neighbors retained by each node during the dynamic graph sparsification process. The complexity of the autoencoder is determined by the full connection calculation of each layer, which is GCN uses sparse matrix multiplication for propagation. The computational complexity of each layer depends on the adjacency matrix |E| and the feature dimension d1. The computational complexity of each layer is O(|E|d l ), the overall complexity is:

[0135] In the dynamic graph learning process, the similarity between all nodes is calculated using cosine similarity, which involves matrix multiplication ZZ T , the computational complexity is: O(N 2 d z ), the calculated similarity matrix N×N scale dense matrix is ​​sparse, and the nearest k top A neighbor.

[0136] This process uses heap sorting, and the complexity is: O(N 2 logk top ), in the clustering and self-supervision modules, the computational complexity of soft assignment is O(NKd z ), the prototype orthogonalization (Gram-Schmidt) complexity is O(K 2 d z ), which are determined by the node-cluster center similarity calculation and orthogonalization operation respectively. Combining the above modules, the overall time complexity of the model is:

[0137]

[0138] Table 1 Clustering process of DynStruct-Cluster algorithm

[0139]

[0140] Based on the present invention, a deep attribute graph clustering method is experimentally verified:

[0141] (1) Benchmark Dataset

[0142] The proposed DynStruct-Cluster is evaluated on five widely used graph benchmark datasets: ACM, DBLP, Citeseer, HHAR, and Reuters. The ACM dataset is used for paper networks, while DBLP is used for author networks. In Citeseer, nodes represent publications, and the connections between nodes represent citation relationships between them. The HHAR dataset contains sensor records from smart devices, and Reuters is a text dataset.

[0143] The ACM dataset contains 3,025 scientific articles covering three major areas: databases, wireless communications, and data mining. Each paper is characterized by an 1,870-dimensional bag-of-words model, and an edge is established between two papers written by the same author. The data comes from conferences such as KDD, SIGMOD, SIGCOMM, and MobiCom.

[0144] Table 2: Statistics of the dataset

[0145]

[0146] (2) Clustering index

[0147] In order to quantitatively evaluate the proposed network DynStruct-Cluster and all baseline methods, four standard and widely used performance indicators are adopted: accuracy ACC, normalized mutual information NMI, average Rand index ARI and macro F1 score F1, where larger values ​​indicate better performance.

[0148] (3) Baseline method

[0149] To verify the clustering performance of the proposed method DynStruct-Cluster on the above five datasets, it will be compared with three types of methods, including clustering methods of original data, DNN-based clustering methods, and GCN-based graph clustering methods.

[0150] K-means: a classic clustering method based on raw data; AE: a two-stage deep clustering algorithm that first learns data representation through an autoencoder and then performs K-means on the representation; DEC: a deep clustering method that designs a clustering objective function to guide the learning of data representation; IDEC: adds reconstruction loss to DEC to learn better representations; GAE&VGAE: an unsupervised graph embedding method that uses GCN (graph convolutional network) to learn data representation; ARGA: proposes a generative adversarial framework to regularize embeddings with prior distributions; DAEGC: uses attention networks to learn Learn node representations and use clustering loss to supervise the graph clustering process; CDRS: proposes a collaborative decision-making-reinforcement self-supervision method to introduce supervisory information to guide the representation learning of graph nodes; SDCN: combines the attribute information learned by AE with the attribute information learned by GCN through the transfer operator to enrich the node embedding, which is an end-to-end graph clustering method; DFCN: introduces an information fusion module to fuse the representations learned by AE and GAE, and trains them through a triple self-supervision strategy; AGCC: This method proposes to update the structure and data representation of the graph layer by layer, uses the attention mechanism to fuse the information of the two networks, and adopts a self-supervision strategy for training.

[0151] (4) Implementation details

[0152] Before training the entire model, the autoencoder (AE) module is pre-trained. The parameters of the pre-training phase are set as follows: the number of training epochs is 50, the learning rate is set to 1e-3, and the batch size is 256;

[0153] On this basis, the complete network is trained end-to-end with 200 training rounds, a random seed of 42, and a uniform layer dimension of AE and GCN of 500-500-2000-10.

[0154] Run the K-Means algorithm 20 times using the pre-trained embedding representation to initialize the cluster centroids, and select the one with the best clustering effect as the final initialization result;

[0155] Each model was run 10 times to reduce the randomness of the clustering results, and the mean and standard deviation of the clustering performance indicators ACC, NMI, ARI, and F1 were reported to improve the stability and reliability of the evaluation.

[0156] The learning rates of various datasets are as follows: for ACM, DBLP, and HHAR, the learning rate is set to 1e-3; for CITE and Reuters, the learning rate is set to 1e-4; for DBLP, the learning rate is set to 1e-3; and for AGCC, the experiments are conducted on the published source code. The settings of the network dimension and learning rate are consistent with those of the deep attribute graph clustering method of the present invention, and the results listed in DFCN are cited.

[0157] All experiments were implemented on a server with a 12vCPU Intel(R)Xeon(R)Platinum 8255C CPU@2.50GHz and an NVIDIA GTX 3090 with 24GB memory.

[0158] (5) Clustering results

[0159] Tables 3 to 7 show the clustering performance of DynStruct-Cluster on five standard datasets (ACM, CiteSeer, DBLP, HHAR, Reuters), and compare it with 12 representative baseline methods on four clustering metrics (ACC, NMI, ARI, F1).

[0160] From the results, it can be seen that DynStruct-Cluster achieves optimal or suboptimal performance on all datasets, especially showing higher stability and robustness on datasets with complex structures or significant heterogeneous modes.

[0161] This demonstrates that the proposed model is able to better capture semantic consistency when learning node representations, and the generated clustering results are highly consistent with the true labels. From an indicator perspective, the steady improvement in ACC and NMI reflects the enhanced overall clustering ability of the model, while the significant improvement in ARI and F1 further highlights the method's stronger discriminative and error correction capabilities when dealing with imbalanced classes and samples with blurred boundaries.

[0162] The specific analysis can be summarized as follows: Compared with methods based solely on attribute information, such as K-Means and Autoencoder, DynStruct-Cluster and structure-aware models such as GAE and DAEGC demonstrated superior clustering results on all datasets. Compared with models using a single GCN structure, DynStruct-Cluster significantly improved clustering performance.

[0163] In summary, DynStruct-Cluster has established a collaborative enhancement mechanism in multiple dimensions, including structural optimization, modal collaboration, and prototype contrastive learning, and achieved stable and superior clustering performance on multiple standard datasets, verifying the effectiveness and versatility of its method design.

[0164] In order to further verify the effectiveness of DynStruct-Cluster, representative multi-network fusion methods such as SDCN and DFCN were selected as comparison objects.

[0165] To ensure a fair comparison, we strictly adhered to the experimental configuration and kept all methods consistent in the hyperparameter settings of the number of iterations, learning rate, optimizer, and activation function in pre-training and formal training. We also ran each method ten times independently and reported the average value of each indicator to improve the stability and credibility of the results.

[0166] like Figure 3 As shown in the figure, experiments show that under the same experimental settings, DynStruct-Cluster achieves the best performance in ACC and ARI indicators on all datasets, and maintains a leading position in NMI and F1 indicators on all datasets except Reuters. Although DFCN is slightly better in NMI and F1 on the Reuters dataset, DynStruct-Cluster still shows a clear advantage in ACC and ARI, demonstrating its superior ability in overall discriminability and global consistency.

[0167] Table 3: Clustering results on the ACM dataset

[0168]

[0169] Table 4: Clustering results on the CITE dataset

[0170]

[0171] Table 5: Clustering results on the DBLP dataset

[0172]

[0173]

[0174] Table 6: Clustering results on the HHAR dataset

[0175]

[0176] Table 7: Clustering results on the Reuters dataset

[0177]

[0178] (6) Parameter analysis

[0179] In order to verify the impact of key hyperparameters in the model on clustering performance, four types of parameters are evaluated: graph structure learning parameters (λ, k top), self-supervised clustering loss weight (λ1), consistency alignment and prototype comparison loss weight (λ2, λ3). Among them, λ is the graph fusion coefficient, which is used to balance the original adjacency matrix and the learned similarity graph; k top The sparsity of the adjacency matrix is ​​controlled to achieve structural denoising by retaining the Top-k most similar edges; λ1 is used to adjust the KL divergence loss strength of the cluster distribution to improve the discriminability of the potential representation; λ2 and λ3 respectively guide node-prototype consistency alignment and inter-class semantic comparison to enhance the distinguishability of cluster boundaries.

[0180] like Figure 4 As shown in the results, it is shown that moderate sparsity helps improve performance - when k top When the number increases from 10 to 15 or 20, the ACC of most datasets is significantly improved; top >20) may destroy the local structure, especially in ACM, which results in performance degradation. top The value is positively correlated with its average node degree, indicating that reasonable sparsity should be adjusted according to the structural characteristics of the original graph.

[0181] In terms of the setting of λ, experiments reveal the heterogeneity of fusion strategies: ACM relies more on learned similar graphs (the optimal λ is relatively low), while CiteSeer tends to balance the fusion strategies. DBLP performs best at a smaller λ due to the dense structure of the original graph.

[0182] Figure 5 Introducing the adaptive λ mechanism, in the optimal k top The adaptive strategy dynamically adjusts the graph fusion ratio under different configurations. Compared to a fixed λ, the adaptive strategy achieves improved ACC on all three datasets, validating the effectiveness of this mechanism in improving performance. Although the final change in λ is small (Δλ ≤ 0.03), the model is still highly sensitive to it, indicating that the graph fusion weight has a significant impact on the overall clustering structure.

[0183] We optimized the loss term ratio, using a heuristics-genetic algorithm strategy to jointly search for λ1, λ2, and λ3 (in the range [0, 1] to [0, 5]). Experiments found that when all three were confined to the range [0, 1], the model clustering performance was optimal, indicating that properly balancing the losses across tasks helps enhance the model's stability and generalization capabilities.

[0184] (7) Ablation experiment

[0185] To verify the effectiveness of the proposed module, three model variants were designed and systematic ablation experiments were conducted on five datasets: ACM, CiteSeer, DBLP, HHAR, and Reuters:

[0186] w / oDGR: remove the dynamic graph structure learning module; w / oDPCL: remove the dynamic prototype comparison module; w / oDGR+DPCL: remove both at the same time, retaining only the AE+GCN module; Full Model: complete model, including DGR and DPCL modules.

[0187] Figure 7 Table 8 shows the performance of each variant on four mainstream clustering metrics (ACC, NMI, ARI, and F1). Overall, the complete model outperforms the model with any missing module on all datasets and metrics, verifying the synergistic contribution of the DGR and DPCL modules.

[0188] In summary, the DGR module is responsible for modeling high-order relationships within the graph structure, while the DPCL module facilitates semantic consistency learning at the prototype level. The two work synergistically to effectively improve clustering performance and training stability. Ablation experiments further validate the effectiveness and necessity of each module design.

[0189] Table 8: Ablation experiment data display

[0190]

[0191]

[0192] Visual analysis based on a deep attribute graph clustering method provided by this invention:

[0193] (1) Visualization of clustering results

[0194] In order to further demonstrate the clustering effect intuitively, the t-SNE algorithm is used to visualize the clustering results of the five data sets in two-dimensional space. Figure 8 It can be seen from FIG that, compared with the irregular distribution of the original data, the model of the present invention makes the samples belonging to the same cluster more concentrated, while the samples belonging to different clusters are further apart.

[0195] (2) Convergence visualization

[0196] Figure 9 The four indicators and loss functions of the five datasets ACM, CITE, DBLP, HHAR, and Reuters are visualized. It is found that with the increase of epochs, the values ​​of the four indicators gradually increase and converge, while the value of the loss function gradually decreases and reaches convergence.

[0197] Therefore, the present invention provides a deep attribute graph clustering method, which improves the discriminability of cluster representation and effectively promotes the adaptive optimization of graph structure.

[0198] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. 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 solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep attribute graph clustering method, characterized in that: The method adopts the dual-drive network framework DynStruct-Cluster, integrates the dual-branch structure of autoencoder AE and graph autoencoder GAE, and clusters attribute graphs by jointly optimizing graph structure and clustering process; the dual-drive network framework DynStruct-Cluster consists of autoencoder AE, graph convolutional network GCN, dynamic graph structure optimization DGR, dynamic prototype contrastive learning DPCL and self-supervised optimization module.

2. A deep attribute graph clustering method according to claim 1, characterized in that: The steps for deep attribute graph clustering using the dual-driver network framework DynStruct-Cluster are: The S1 and DGR modules dynamically optimize the structure of the original adjacency matrix and update the adjacency matrix according to the current node features and the structure of the graph in each training iteration. The input data consists of the node feature matrix X and the adjacency matrix A. S2, AE and GCN modules learn two heterogeneous representations from the attributes and structure of the graph respectively. Each layer of features of the AE encoder is fused with each layer of features of the GCN to obtain the final fused features; S3, perform prototype comparison through the DPCL module and calculate the contrast loss; S4. Generate soft assignments Q and Q′ of clusters through the latent representation and the final fused features, where the target distribution p is generated by Q and guides the self-supervised clustering process.

3. A deep attribute graph clustering method according to claim 1, characterized in that: The autoencoder AE module learns low-dimensional representations from node attributes and generates a potential representation Z to assist downstream node clustering tasks; the autoencoder AE adopts a symmetrical structure and consists of two parts: an encoder and a decoder. The encoder maps the original data to a low-dimensional latent space by stacking multiple layers of nonlinear transformations, and the decoder reconstructs the original feature data to retain the main structural information of the input data.

4. A deep attribute graph clustering method according to claim 3, characterized in that: The specific process of encoding and decoding by the autoencoder AE and performing optimization training to generate the potential representation Z is as follows: Step 1: In the encoding stage, the original input data is Where N is the number of samples, d is the input feature dimension, and assuming that AE consists of L layers, the hidden representation of the lth layer is: Where W e (l) and denote the weight matrix and bias vector of the corresponding layer respectively, σ(·) is the activation function ReLU of the fully connected layer; Step 2: In the decoding stage, the decoder reconstructs the input layer by layer, and the decoder output formula is: The final output is the reconstruction result Step 3: The model is trained by minimizing the difference between the input and the reconstruction. The optimization goal is: Where x i and Represent the original features and reconstructed features of the i-th sample, respectively, ‖·‖ F Represents the Frobenius norm, and finally the potential representation Z obtained by the last layer output of the encoder, where Z = Y (L) Serves as input for subsequent graph structure learning and clustering optimization.

5. A deep attribute graph clustering method according to claim 1, characterized in that: The graph convolutional network (GCN) module captures the topological relationship between nodes and their neighbors and fuses the topological information of GCN with the feature representation of AE.

6. A deep attribute graph clustering method according to claim 5, characterized in that: The specific process of fusing the topological information of GCN with the feature representation of AE is as follows: a linear weighted fusion strategy is adopted to unify the structural information and attribute information into a unified model. The calculation expression is: Where, is the output of the AE layer l, α = 0.5; Combine AE and GCN layer by layer to obtain a new fusion feature representation. The calculation expression is: Where, represents the adjacency matrix after adding the self-loop, Its degree matrix, W (l) is the weight matrix of the lth layer, σ is the nonlinear activation function; The last layer of GCN output is passed through Softmax to generate a cluster probability distribution, and a multi-category probability distribution matrix is ​​obtained, in which each row is a probability distribution of a sample. The calculation formula is: Q′=soft max(H (L) ); Where Q′ represents each element Q′ ij Represents the probability that sample i belongs to category j.

7. A deep attribute graph clustering method according to claim 1, characterized in that: The dynamic graph structure optimization module DGR dynamically optimizes the graph structure through three stages: dynamic similarity modeling, adaptive sparsification, and graph structure fusion. The dynamic similarity modeling stage calculates the dynamic similarity between nodes based on node attributes and structure information. The calculation formula is: Where, X i represents the input feature of the i-th node, X is the original feature of the current node, and S is the dynamic similarity graph; The adaptive sparsification stage performs a row-by-row Top-k screening on the similarity matrix, retaining the most relevant K neighbors of each node while excluding self-loops, generating a sparse similarity graph, and symmetric normalizing the sparse similarity graph to construct an optimized adjacency matrix. The calculation formula is: Where, is a sparse similarity graph, A s is the adjacency matrix, D is the degree matrix, and its diagonal elements are defined as The graph structure fusion stage fuses the optimized graph structure with the original graph structure to obtain the final graph structure and participate in the feature propagation process of GCN. The optimized propagation form is expressed as: in, is the final adjacency matrix, α∈[0,1] is the learnable fusion coefficient.

8. A deep attribute graph clustering method according to claim 1, characterized in that: The dynamic prototype contrastive learning module (DPCL) uses a prototype clustering mechanism to enhance the discriminability of sample representations, optimizes prototype representations through a momentum update strategy, and enhances intra-class consistency and inter-class discrimination. The prototype clustering mechanism uses the central representation of each category as the prototype and performs cluster assignment by calculating the distance between the sample and the prototype. The momentum update strategy dynamically updates the prototype representation based on historical prototype representations and current sample representations.

9. A deep attribute graph clustering method according to claim 8, characterized in that: The specific process of forming a clustering structure using the dynamic prototype contrastive learning module DPCL is as follows: First, the K-means algorithm is used to cluster the high-order feature representation based on the GCN output to generate preliminary pseudo labels. Based on these pseudo labels, the prototype vector of each class is calculated: Where C k is the prototype vector of each class, Represents the set of nodes currently assigned pseudo labels as the Kth class, V k is the number of nodes in the set, h i Represents the feature representation of node i, which is the final fusion representation of the GCN module; Secondly, the momentum update strategy is used to dynamically adjust the prototype vector of each class. The update formula is as follows: Where β∈[0,1] is a fixed momentum coefficient; Finally, maximize the similarity of each sample with its prototype and minimize its similarity with prototypes of other categories, given the temperature parameter τ , calculate the similarity matrix between all nodes and the prototype, and define the prototype contrast loss of each sample as the normalized cross entropy to form a more discriminative clustering structure. The calculation formula is:

10. A deep attribute graph clustering method according to claim 1, characterized in that: The self-supervisory optimization module optimizes model parameters through self-supervisory signals; the self-supervisory signals include but are not limited to feature reconstruction loss, cluster optimization loss, cluster distribution consistency constraint and dynamic prototype comparison loss; the feature reconstruction loss, cluster optimization loss, cluster distribution consistency constraint and dynamic prototype comparison loss constitute the total loss function, which is expressed as: <h2 style=";text-align:left;direction:ltr">L=L<h2 style=";text-align:left;direction:ltr"> res <h2 style=";text-align:left;direction:ltr"> +<h2 style=";text-align:left;direction:ltr"> λ <h2 style=";text-align:left;direction:ltr"> 1L<h2 style=";text-align:left;direction:ltr"> clu <h2 style=";text-align:left;direction:ltr"> +<h2 style=";text-align:left;direction:ltr"> λ <h2 style=";text-align:left;direction:ltr"> 2L<h2 style=";text-align:left;direction:ltr"> cess <h2 style=";text-align:left;direction:ltr"> +<h2 style=";text-align:left;direction:ltr"> λ <h2 style=";text-align:left;direction:ltr"> 3L<h2 style=";text-align:left;direction:ltr"> proto <h2 style=";text-align:left;direction:ltr"> ; Where, L res is the reconstruction loss of the feature, L clu Optimize the loss for clustering, L cess is the cluster distribution consistency constraint, L DPCL is the dynamic prototype contrast loss, and λ1, λ2, and λ3 are hyperparameters.

Images