Adaptive graph clustering method and system based on multi-view contrast reinforcement learning
Through the adaptive graph clustering method of multi-view comparison reinforcement learning, the multi-view data fusion and graph structure complex adaptability problems are solved, and the accuracy and stability of graph clustering are improved. It is suitable for social networks, recommendation systems, bioinformatics and financial risk control fields.
Patent Information
- Application Number
- CN202510358537.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-25
- Publication Date
- 2025-07-11
AI Technical Summary
When processing multi-view data, the existing graph clustering methods are difficult to effectively integrate information from different views, and cannot fully mine and utilize multi-view information. They are not adaptable when facing complex graph structures, and noise and structural changes affect clustering stability.
Adaptive graph clustering method based on multi-view contrast reinforcement learning is adopted, and feature enhancement is performed by obtaining the adjacency matrix of multi-views, local information matrix and global information matrix are generated, and local and global information matrix are aggregated, perturbations are applied for comparison learning, and finally the clustering results are adjusted through reinforcement learning.
It significantly improves the feature expression ability of graph data and the accuracy of clustering results, enhances the robustness and adaptability of the model, can dynamically adjust the clustering boundaries, reduce the influence of noise and outliers, and improves the accuracy and interpretability of clustering.
Smart Images
Figure CN120296457A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of multi-view image clustering, and relates to an adaptive graph clustering method and system based on multi-view contrast reinforcement learning. Background Art
[0002] Graph clustering, as a key task in the fields of data mining and machine learning, has shown extensive application value in many fields such as social network analysis, recommendation system construction, and disease spread prediction. In recent years, with the rapid development of deep learning technologies such as graph convolutional neural networks (GCNs), graph-based clustering methods have made remarkable progress, especially in node representation learning and graph structure feature capture.
[0003] GCNs perform node representation learning by aggregating the neighborhood information of nodes, and this mechanism enables them to effectively capture the local features of the graph structure, providing strong support for the graph clustering task. However, although GCNs have achieved many accomplishments in the field of graph clustering, they still face various challenges in practical applications, especially in dealing with multi-view data, unified modeling of feature expressions, and adaptability to graph structure complexity.
[0004] Graph data usually has multi-view characteristics, that is, the same node may exhibit different features and similarity metrics under different views. Taking a social network as an example, the similarity of nodes can be measured from multiple perspectives such as the user behavior view and the social relationship view. However, traditional GCN methods are often limited to single-view learning and are difficult to effectively process complex and diverse multi-view data. This results in poor performance of the model when facing diverse data and fails to fully exploit and utilize the advantages brought by multi-view information.
[0005] To overcome this limitation, multi-view representation learning methods have emerged. These methods improve the performance of the model by simultaneously using multiple view information. However, how to effectively fuse the information of different views remains a technical challenge. Existing multi-view learning methods, such as "Graph Convolutional Networks with Adaptive Neighborhood Awareness" (MVGCN) published by Guang et al. in IEEE Transactions on Pattern Analysis and Machine Intelligence, although to a certain extent achieve independent modeling of graphs with multiple views and subsequent view fusion, their fusion methods mostly use simple weighted averaging and are difficult to fully capture the complex relationships and complementarities between different views.
[0006] In addition, the neighborhood awareness methods in GCN also have limitations. Traditional neighborhood awareness algorithms usually adopt static graph structures or neighborhood aggregation strategies, failing to consider the complementarity of information at different levels and different hop counts in the graph structure. This results in insufficient ability of the model to adapt to complex graph structures and difficulty in dynamically adjusting the weights between features at different levels. Therefore, how to design a neighborhood awareness method that can dynamically adjust the weights of information at each layer has become the key to improving the model's adaptability and clustering effect.
[0007] Finally, in multi-view learning, noise and structural changes are also important factors affecting the stability of clustering. Existing multi-view learning methods often lack effective screening and optimization mechanisms when dealing with noisy nodes and edge nodes, resulting in the possible inclusion of irrelevant or unimportant nodes in the clustering results, thus affecting the clustering effect and stability.
[0008] In summary, although deep learning technologies such as GCN provide powerful tools for graph clustering, they still face various challenges in practical applications, such as multi-view data processing, unified modeling of feature expressions, adaptability to graph structure complexity, and handling of noise and structural changes. Summary of the Invention
[0009] The purpose of the present invention is to solve the technical problems of unstable and inaccurate graph clustering effect in the prior art, and provide an adaptive graph clustering method and system based on multi-view contrast reinforcement learning.
[0010] To achieve the above object, the present invention adopts the following technical solutions: The first aspect of the present invention provides an adaptive graph clustering method based on multi-view contrast reinforcement learning, including the following steps: Obtain the adjacency matrices of multiple views, perform feature enhancement on the adjacency matrices of multiple views to generate local information matrices and global information matrices; Aggregate the node information of the local information matrices and the global information matrices to obtain the aggregated representations of the local information matrices and the aggregated representations of the global information matrices; Respectively apply perturbations to the aggregated representations of the local information matrices and the aggregated representations of the global information matrices, and perform contrastive learning on the perturbed aggregated representations of the local information matrices and the aggregated representations of the global information matrices to obtain the optimal local information features and the optimal global information features; Aggregate the optimal local information features and the optimal global information features to obtain an optimized node representation matrix; Perform clustering mapping on the optimized node representation matrix to obtain a preliminary clustering result; Adjust the preliminary clustering result to obtain the final clustering result.
[0011] Furthermore, the adjacency matrix of the multi-views is enhanced in features to generate a local information matrix and a global information matrix, specifically as follows: The local information matrix is obtained by enhancing the adjacency matrix ; The local information matrix is described as:
[0012] where represents the feature of node , represents the degree of node , is set according to the degree of the node, represents the feature of the neighbor node; The global structural feature of the node is extracted by Laplacian matrix decomposition to generate the global information matrix .
[0013] Furthermore, the node information of the local information matrix and the global information matrix is aggregated, specifically as follows: Based on the GCN network, for the local information matrix and the global information matrix, the node features of each layer of the GCN network and the corresponding neighbor node features are weighted and aggregated to obtain the node representation of each layer; calculate the importance score of the node representation of each layer; update the node representation of each layer according to the importance score; The weighted sum of the updated node representations of different channels is calculated to obtain the aggregated representation of the local information matrix and the aggregated representation of the global information matrix.
[0014] Furthermore, the aggregation of the node features of each layer of the GCN network and the corresponding neighbor node features is specifically as follows:
[0015] where represents the node feature representation matrix of the k th layer, represents the weight matrix of the k th layer, represents the non-linear activation function, is the normalized adjacency matrix, is the node feature representation matrix of the k+ 1st layer.
[0016] Furthermore, perturbations are respectively applied to the aggregated representation of the local information matrix and the aggregated representation of the global information matrix, specifically as follows:
[0017]
[0018] Among them, represents the aggregated representation of the local information matrix or the aggregated representation of the global information matrix, represents the noise matrix, represents the node i features, represents the node i features after adding noise.
[0019] Furthermore, the contrastive learning of the aggregated representation of the perturbed local information matrix and the aggregated representation of the global information matrix is as follows: For the same view, maximize the aggregated representation of the local information matrix and the aggregated representation of the global information matrix and the aggregated representation of the perturbed local information matrix and the aggregated representation of the global information matrix; For different views, maximize the information of the same node under different views; The loss function of the contrastive learning is:
[0020] Among them, the loss function of the contrastive learning for the same view is:
[0021] The loss function of the contrastive learning for different views is:
[0022] In the formula, is the temperature parameter, used to control the smoothness of the distribution, represents the node representation of the perturbed view, represents the similarity between node representations; represents the local node representation; represents the global node representation; represents the weight balance factor, used to control and the relative importance of. Furthermore, adjust the preliminary clustering result to obtain the final clustering result, specifically: Decompose the preliminary clustering result into several clustering clusters; perform node addition or node deletion operations for each clustering cluster; After each node addition or node deletion operation, calculate the reward for the corresponding operation; When the reward no longer increases significantly or reaches the maximum number of iterations, stop the operation to obtain the final clustering result.
[0023] Furthermore, the multi-view includes the user's social relationship network, interest preferences, and behavior pattern information..
[0024] In a second aspect of the present invention, there is provided a computer-readable storage medium storing a computer program which, when executed by a processor, implements the above-mentioned adaptive graph clustering method based on multi-view contrastive reinforcement learning.
[0025] In a third aspect of the present invention, there is provided an adaptive graph clustering system based on multi-view contrastive reinforcement learning, comprising: A feature generation module, which obtains the adjacency matrices of multiple views, enhances the features of the adjacency matrices of multiple views, and generates a local information matrix and a global information matrix; A feature aggregation module, which aggregates the node information of the local information matrix and the global information matrix to obtain the aggregated representations of the local information matrix and the global information matrix; An optimal feature learning module, which respectively perturbs the aggregated representation of the local information matrix and the aggregated representation of the global information matrix, and performs contrastive learning on the perturbed aggregated representation of the local information matrix and the aggregated representation of the global information matrix to obtain the optimal local information features and the optimal global information features; A feature optimization module, which aggregates the optimal local information features and the optimal global information features to obtain an optimized node representation matrix; A clustering mapping module, which performs clustering mapping on the optimized node representation matrix to obtain a preliminary clustering result; A clustering adjustment module, which adjusts the preliminary clustering result to obtain a final clustering result.
[0026] Compared with the prior art, the present invention has the following beneficial effects: The present invention discloses an adaptive graph clustering method based on multi-view contrastive reinforcement learning. By obtaining the adjacency matrices of multiple views and performing feature enhancement, this method can make full use of the complementary information between different views in the graph data to generate a local information matrix and a global information matrix that contain both local details and global structures. This step significantly improves the feature expression ability of the graph data and lays a solid foundation for subsequent clustering tasks. By aggregating the node information of the local information matrix and the global information matrix, respective aggregated representations are obtained. This method effectively integrates the local connection information between nodes and the global topological structure, improving the efficiency and accuracy of information aggregation. This aggregation strategy helps to reveal the deep relationships between nodes and provides a richer and more accurate information basis for contrastive learning. By applying perturbations to the aggregated local information features and global information features and performing contrastive learning, this method can identify and strengthen the node features that can remain stable under different perturbations, thereby extracting the optimal local information features and global information features. This process not only enhances the robustness of the model but also significantly improves the accuracy of the clustering results. Aggregating the optimal local information features and the optimal global information features yields an optimized node representation matrix, which more accurately reflects the true positions and roles of nodes in the graph. Based on this optimized node representation for clustering mapping, a more precise and stable preliminary clustering result can be obtained. By adjusting the preliminary clustering result, this method can further refine the clustering boundary, reduce the influence of noise and outliers, and thus obtain a final clustering result that more conforms to the true distribution of the data. This adjustment mechanism enhances the adaptability and flexibility of the method, enabling it to be widely applied to different types of graph data and clustering requirements.
[0027] Furthermore, by performing detailed node addition or deletion operations on the preliminary clustering result, this method can dynamically adjust the boundaries of the clustering clusters, reduce the interference of noise and outliers, and thus significantly improve the clustering accuracy. This adjustment mechanism ensures that the clustering result is closer to the true distribution of the data, improving the accuracy and reliability of the clustering. The reinforcement learning mechanism enables the model to learn the optimal clustering strategy through continuous trial and adjustment. Even in the face of complex and changing graph data and different clustering requirements, this method can maintain stable clustering performance through self-learning and optimization. The reward calculation process in the reinforcement learning mechanism provides a clear basis for the optimization of the clustering result. By checking which operations have obtained higher rewards, it is easier to understand the formation process of the clustering result, thereby improving the interpretability of the clustering result. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.
[0029] Figure 1 It is a flowchart of the adaptive graph clustering method based on multi-view contrast reinforcement learning of the present invention; Figure 2 It is a process diagram of the adaptive graph clustering method based on multi-view contrast reinforcement learning of the present invention. Specific embodiments
[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and marked in the accompanying drawings here can be arranged and designed in various different configurations.
[0031] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed present invention, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.
[0032] It should be noted that similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.
[0033] The following will further describe the present invention in detail with reference to the accompanying drawings: See Figure 1 , the present invention provides an adaptive graph clustering method based on multi-view contrast reinforcement learning, including the following steps: S1: Obtain the adjacency matrices of multiple views, perform feature enhancement on the adjacency matrices of multiple views, and generate a local information matrix and a global information matrix; Specifically, for a certain node v , without considering the node attribute information, perform feature enhancement on it: .
[0034] Among them, x'( v ) represents the feature of node v , D( v) represents a node v 's degree, and num( v ) represents setting according to the degree of the node. When the degree is greater than 5, it is 1, otherwise it is 0. B( v ) represents the node neighbors.
[0035] Subsequently, spectral features are extracted through the Laplacian matrix to obtain a feature matrix centered on global information. The normalized graph Laplacian matrix is obtained through the following formula L :
[0036] where I is the identity matrix, A is the adjacency matrix of the graph, and D is the degree matrix. Next, L is eigen-decomposed through the following formula to obtain eigenvalues and eigenvectors:
[0037] where represents the eigenvector matrix, and the column vector is the eigenvector, represents the eigenvalue diagonal matrix. Subsequently, the first k eigenvectors are selected to form the matrix : .
[0038] The global feature of each node is composed of the components of the corresponding k eigenvectors.
[0039] A dual view is constructed through the above method. The local view (local information matrix) reflects the neighborhood characteristics, and the global view (global information matrix) reflects the macroscopic position. The combination of the two can capture the graph structure characteristics more comprehensively.
[0040] S2: Aggregate the node information of the local information matrix and the global information matrix to obtain the aggregated representation of the local information matrix and the aggregated representation of the global information matrix; specifically: S201, after generating the local information matrix and the global information matrix, based on GCN, by aggregating the node information of each layer of the GCN network and the information of its neighbors, further improve the ability to capture the characteristics of the local and global graph structures. Through the information propagation of each layer of the network, GCN aggregates the features of nodes and their neighbors in the local view or the global view, captures a larger context range, and thus learns node representations. Specifically as follows:
[0041] where represents the node feature representation matrix of the k th layer, represents the weight matrix of the k th layer, represents the non-linear activation function, where is the original feature matrix.
[0042] S202, traverse the node representations of each layer of the GCN network and calculate a selection factor . The selection factor represents the importance of the layer, and the larger the value, the more important it is. When the selection factor is 0, it means that the corresponding layer will not be selected. For the initial aggregated representation:
[0043] where, is the node representation of the first layer. Then the aggregated representation is concatenated with the node representation of the next layer, and a linear transformation is performed on the concatenated features to obtain the corresponding associated representation:
[0044]
[0045] where, represents the concatenation operation, N is the number of nodes, d is the feature dimension, W is the learnable parameter matrix. represents the weighted node representation of the th layer. Next, the importance scores of and are calculated respectively through the following formulas:
[0046]
[0047] where, is the learnable parameter vector. and represent the importance scores of and respectively, and selection is performed through the importance scores. Subsequently, the scores are mapped to the interval [0, 1] by applying an activation function:
[0048]
[0049]
[0050] where, is the Sigmoid activation function. The aggregated representation is selectively updated according to the final importance score to achieve weighting of the representations of each layer.
[0051] S203, aggregate the node representations of different channels using the attention mechanism. Calculate the importance scores of each view, and use weights to perform weighted summation on the view representations to generate the final node representation, thereby obtaining the aggregated representation of the local information matrix and the aggregated representation of the global information matrix:
[0052] S3: Apply perturbations to the aggregated representation of the local information matrix and the aggregated representation of the global information matrix respectively, and perform contrastive learning on the aggregated representation of the perturbed local information matrix and the aggregated representation of the global information matrix to obtain the optimal local information features and the optimal global information features; specifically: S301, for the learned feature matrix , add a certain perturbation to it, specifically as follows:
[0053] Specifically, the noise matrix is described as: ; where, represents the aggregated representation of the local information matrix or the aggregated representation of the global information matrix, represents the noise matrix, which is used to perturb the aggregated representation, represents the feature of node i , represents the feature of node i after adding noise.
[0054] S302, use the loss function of contrastive learning to maximize the similarity between the original view (aggregated representation) and the perturbed view (aggregated representation after adding perturbation). Define the node representation matrices of the local view and the global view as and respectively, and the specific formula is as follows:
[0055] where, is the temperature parameter, which is used to control the smoothness of the distribution, represents the representation of node i in the local view, represents the node representation of the perturbed view, is the similarity function, which represents the similarity between node representations. By maximizing the similarity between the same node in the local view and the perturbed representation, the node representation can be made more robust. Among them, is used to minimize redundant information. By optimizing the relationship between different node representations, the redundancy of the node representation is reduced, ensuring that the generated node representation is more discriminative. It is used to maximize the mutual information of positive sample pairs, ensure that the representations of the same node in different views are as similar as possible, and enhance the aggregation and consistency of the representations.
[0056] S303. For contrastive learning between different views, maximize the mutual information of the same node in different views and enhance the aggregation and consistency of the representations. The specific formula is as follows:
[0057] where, represents the node i 's representation in the global view. By comparing the node representations in different views, the model can simultaneously consider the local features and global structure information of the node and generate a more comprehensive and rich node representation. Finally, the loss function of the contrastive learning part can be expressed by the formula:
[0058] where, The purpose is to contrast the same node in the local view and the perturbed representation. It is used to maximize the mutual information of positive sample pairs to ensure that the representations of the same node in different views are as similar as possible. is the weight balancing factor used to control and 's relative importance. The final node representation is obtained through this method. Subsequently, through for clustering mapping to obtain the preliminary clustering result .
[0059] S4: Adjust the preliminary clustering result to obtain the final clustering result. Specifically: Divide the final clustering result into multiple clustering clusters; For each clustering cluster, perform fine-tuning, specifically: S401. Train a rewriter to fine-tune the clustering cluster through a reinforcement learning framework. The rewriter can perform addition or removal operations on the nodes in the current state and calculate the rewards for the corresponding operations. If the reward is positive, it means that the current operation is beneficial to the clustering result and the current operation is executed; otherwise, it means that the current operation is not beneficial to the clustering result and the current operation is not executed. Subsequently, update the parameters of the rewriter model through the policy gradient method so that it can better select high-quality actions in future rewriting tasks and dynamically update the node representations at the same time.
[0060] For the node v in the preliminary clustering result, at time , its state is as follows:
[0061] where, Represents at time t the state of node v , where represents the node representation obtained in the previous step in the characteristics of node
[0062] For the addition operation, the node representation of the cluster is updated by combining the neighborhood node information of the cluster through GNN. For the removal operation, the edge nodes of the current cluster are removed. The specific formulas are as follows:
[0063]
[0064] S402, through the operation to select an action and modify the node characteristics in the current state, represents performing a removal operation on the current node, represents performing an addition operation on the current node, represents at time t the cluster at time represents the cluster the set of neighbor nodes of
[0065] After performing a certain action, calculate the reward brought before and after the execution of this action , and the specific formula is as follows:
[0066] Among them, represents the score calculation function, which represents the initial clustering result. Judge the correctness of the current operation through the positive and negative of the reward
[0067] S403, update the parameters of the rewriter network through the proximal policy optimization method, as shown in the specific formula:
[0068] Among them, is the ratio of the current policy to the old policy, is the advantage function, which represents the advantage degree of performing a certain action, represents the hyperparameter for limiting the amplitude of policy update.
[0069] Finally, update the parameters of the rewriter network according to the objective function :
[0070] Among them, is the learning rate, represents the gradient of the loss function. The PPO algorithm is used to optimize the action policy of the node, enabling the rewriter network to select the optimal action according to the current state, such as node addition and deletion operations. S404, set the maximum number of iterations M , if the reward no longer increases significantly or the maximum number of iterations is reached, stop training to obtain the final rewriter.
[0071] S405: The current number of iterations is less than M , update the parameters of the rewriter model through gradient descent.
[0072] S406: Adjust the initial clustering through the rewriter to obtain the final clustering result.
[0073] The adaptive graph clustering method based on multi-view contrast reinforcement learning proposed by the present invention integrates multi-view contrast learning and reinforcement learning strategies, realizing a clustering process from rough to fine, aiming to significantly improve the accuracy, robustness, and stability of graph clustering. The method of the present invention is applicable to the following multiple fields: Social network analysis: As typical graph-structured data, social networks contain intricate relationships and interaction information among users. Through multi-view perspectives, such as users' social networks, interest preferences, and behavior patterns, user characteristics can be captured more comprehensively and multi-angularly, thus optimizing the clustering effect.
[0074] Recommendation systems: In the graph structure composed of users and items, the reinforcement learning module can more precisely group users or items by dynamically adjusting the clustering structure, which not only improves the accuracy of recommendation results but also enhances the personalized recommendation experience.
[0075] Bioinformatics: In biological networks, elements such as genes, proteins, and diseases are interconnected through graph structures. The multi-view graph clustering method can integrate multi-source information such as the structure, expression data, and functional annotations of genes, providing a more comprehensive perspective for disease correlation analysis. For example, in the study of cancer-related genes, by constructing a gene-disease graph and integrating various information, more in-depth clustering analysis can be achieved.
[0076] Financial risk control: In the financial field, risk signals are hidden in the complex graph structure composed of customers, transactions, accounts, etc. The multi-view graph clustering method effectively identifies the hidden associations between customers and abnormal transaction patterns by integrating multi-dimensional information such as transaction records, customer backgrounds, and account behaviors, which is of great significance for anti-money laundering and financial risk control.
[0077] To further illustrate the superiority of the method of the present invention, the present invention uses three indicators, namely F1 Score, Jaccard Score, and NMI, as evaluation indicators and compares them with six advanced algorithms. The specific experimental results are shown in the table:
[0078] The present invention has achieved significant performance improvement on multiple benchmark datasets. From the table, it can be found that: In the main evaluation indicators (F1, NMI, Jaccard), the present invention is superior to the compared algorithms on all datasets. Under large-scale datasets, it shows higher stability and adaptability. On the Amazon dataset, the F1 Score of the present invention reaches 0.796, which is 0.022 higher than 0.779 of CMGEC, and the NMI reaches 0.754, which is 0.048 higher than the optimal comparison method. These results fully demonstrate the effectiveness of the present invention in multi-view clustering tasks. At the same time, on the LiveJournal dataset, CLARE performs better than CMGEC. The main reason for the analysis is that the subgraph structure features of the LiveJournal dataset are relatively clear and the noise is relatively small, which is suitable for the generative adversarial network optimization model used by CLARE. On the Amaon and DBLP datasets, the graph embedding consistency method proposed by CMGEC can better integrate different views.
[0079] In the Amazon and dblp datasets, the node degree distribution shows a power-law distribution. The connection degrees of most products are relatively low, and only a few products have high connection degrees. Therefore, overall, the overall density of the graph is low, but dense subgraphs will be formed in some local areas. The performance of the present invention on the Amazon dataset is significantly higher than that of other comparison algorithms, mainly because the present invention adopts an adaptive hybrid neighborhood perception mechanism, which can dynamically adjust the perception range of the multi-hop neighborhood, capture the information of distant nodes by expanding the hop count in a sparse network, so as to make up for the defect of insufficient local neighborhood information.
[0080] CMGE through cross-view Figure 1Consistent embedding for multi-view information integration can utilize the correlations between multi-views. However, its method is overly dependent on fixed graph embeddings and consistency constraints, making it difficult to flexibly adapt to graph structure heterogeneity and weight differences between views. In contrast, the adaptive hybrid neighborhood-aware GCN proposed in this invention can dynamically adjust the importance of features at each layer, enabling the adaptive fusion of local and global information and effectively capturing complex structural relationships in the graph. Additionally, this invention fine-tunes the clustering results through a reinforcement learning module, which can dynamically optimize node allocation and community structure, enhancing the ability to handle noisy nodes and marginal nodes, thereby achieving higher clustering accuracy and stability on complex and heterogeneous datasets. In comparison, CMGE performs poorly in noise handling and adaptability to heterogeneous data and is vulnerable to data noise and view inconsistency.
[0081] An embodiment of this invention provides an adaptive graph clustering system based on multi-view contrastive reinforcement learning, including: A feature generation module that obtains the adjacency matrices of multiple views, enhances the features of the adjacency matrices of multiple views, and generates a local information matrix and a global information matrix; A feature aggregation module that aggregates the node information of the local information matrix and the global information matrix to obtain the aggregated representation of the local information matrix and the aggregated representation of the global information matrix; An optimal feature learning module that respectively perturbs the aggregated representation of the local information matrix and the aggregated representation of the global information matrix, performs contrastive learning on the perturbed aggregated representation of the local information matrix and the aggregated representation of the global information matrix, and obtains the optimal local information features and the optimal global information features; A feature optimization module that aggregates the optimal local information features and the optimal global information features to obtain an optimized node representation matrix; A clustering mapping module that performs clustering mapping on the optimized node representation matrix to obtain a preliminary clustering result; A clustering adjustment module that adjusts the preliminary clustering result to obtain the final clustering result.
[0082] In another embodiment of the present invention, the present invention further provides a storage medium, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in a terminal device, used to store programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and, of course, the extended storage medium supported by the terminal device. It can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution system, device, or component. The computer-readable storage medium provides storage space, and this storage space stores the operating system of the terminal. And, in this storage space, one or more instructions suitable for being loaded and executed by the processor are also stored. These instructions can be one or more computer programs (including program codes). It should be noted that more specific examples (non-exhaustive list) of the computer-readable storage medium here include: electrical connections with one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.
[0083] The computer-readable storage medium also includes data signals propagated in a baseband or as part of a carrier wave, in which the readable program code is carried. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The readable storage medium can also be any readable medium other than the readable storage medium, and this readable medium can send, propagate, or transmit a program for use by or in combination with an instruction execution system, device, or component. The program code contained on the readable storage medium can be transmitted by any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination of the above.
[0084] The program code for performing the operations of the present invention can be written in any combination of one or more programming languages. The programming languages include object-oriented programming languages - such as Java, C++, etc., and also include conventional procedural programming languages - such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, executed as an independent software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the user's computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computing device (for example, by using an Internet service provider to connect through the Internet).
[0085] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the adaptive graph clustering method based on multi-view contrastive reinforcement learning in the above embodiments.
[0086] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.
Claims
1. An adaptive graph clustering method based on multi-view contrastive reinforcement learning, characterized in that, It includes the following steps: Obtain the adjacency matrix of multiple views, perform feature enhancement on the adjacency matrix of multiple views, and generate a local information matrix and a global information matrix; Aggregate the node information of the local information matrix and the global information matrix to obtain the aggregated representation of the local information matrix and the aggregated representation of the global information matrix; Apply perturbations to the aggregated representation of the local information matrix and the aggregated representation of the global information matrix respectively, and perform contrastive learning on the aggregated representation of the local information matrix and the aggregated representation of the global information matrix after perturbation to obtain the optimal local information features and the optimal global information features; Aggregate the optimal local information features and the optimal global information features to obtain an optimized node representation matrix; Perform clustering mapping on the optimized node representation matrix to obtain a preliminary clustering result; Adjust the preliminary clustering result to obtain the final clustering result.
2. The adaptive graph clustering method based on multi-view contrastive reinforcement learning according to claim 1, wherein The feature enhancement of the adjacency matrix of multiple views to generate a local information matrix and a global information matrix is specifically as follows: By enhancing the adjacency matrix, a local information matrix is obtained ; The local information matrix is described as follows: Among them, represents the characteristics of the node ; represents the degree of the node ; is set according to the degree of the node, represents the characteristics of neighbor nodes; Extract the global structural features of nodes through Laplacian matrix factorization to generate a global information matrix .
3. The adaptive graph clustering method based on multi-view contrastive reinforcement learning according to claim 1, characterized in that, The aggregation of the node information of the local information matrix and the global information matrix is specifically as follows: Based on the GCN network, for the local information matrix and the global information matrix, weightedly aggregate the node features of each layer of the GCN network and the corresponding neighbor node features to obtain the node representation of each layer; calculate the importance score of the node representation of each layer; update the node representation of each layer according to the importance score; Perform weighted summation on the updated node representations of different channels to obtain the aggregated representation of the local information matrix and the aggregated representation of the global information matrix.
4. The adaptive graph clustering method based on multi-view contrastive reinforcement learning according to claim 3, wherein The aggregation of the node features of each layer of the GCN network and the corresponding neighbor node features is specifically as follows: Among them, represents the node feature representation matrix of the k th layer, represents the weight matrix of the k th layer, represents the non-linear activation function, is the normalized adjacency matrix, is the node feature representation matrix of the k+ 1st layer.
5. The adaptive graph clustering method based on multi-view contrastive reinforcement learning according to claim 1, wherein The perturbations applied to the aggregated representation of the local information matrix and the aggregated representation of the global information matrix respectively are specifically as follows: Among them, represents the aggregated representation of the local information matrix or the aggregated representation of the global information matrix, represents the noise matrix, represents the node i features, represents the node i features after adding noise.
6. The adaptive graph clustering method based on multi-view contrastive reinforcement learning according to claim 1, wherein The contrastive learning of the aggregated representation of the local information matrix and the aggregated representation of the global information matrix after perturbation is specifically as follows: For the same view, maximize the aggregated representation of the local information matrix and the aggregated representation of the global information matrix and the aggregated representation of the local information matrix and the aggregated representation of the global information matrix after perturbation; For different views, maximize the information of the same node under different views; The loss function of the contrastive learning is: Among them, the loss function of the contrastive learning for the same view is: The loss function of the contrastive learning for different views is: In the formula, is the temperature parameter, which is used to control the smoothness of the distribution, represents the node representation of the perturbed view, represents the similarity between node representations; represents the local node representation; represents the global node representation; represents the weight balance factor, which is used to control and the relative importance of.
7. The adaptive graph clustering method based on multi-view contrastive reinforcement learning according to claim 1, wherein Adjust the preliminary clustering result to obtain the final clustering result, specifically as follows: Decompose the preliminary clustering result into several clustering clusters; perform node addition or node deletion operations on each clustering cluster; For each node addition or node deletion operation completed, calculate the reward corresponding to the operation; When the reward no longer increases significantly or reaches the maximum number of iterations, stop the operation to obtain the final clustering result.
8. The adaptive graph clustering method based on multi-view contrastive reinforcement learning according to claim 1, characterized in that The multiple views include the user's social relationship network, interest preferences, and behavior pattern information.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the adaptive graph clustering method based on multi-view contrastive reinforcement learning according to any one of claims 1-7.
10. An adaptive graph clustering system based on multi-view contrastive reinforcement learning, characterized in that It includes: A feature generation module that obtains the adjacency matrices of multiple views, enhances the features of the adjacency matrices of multiple views, and generates a local information matrix and a global information matrix; A feature aggregation module that aggregates the node information of the local information matrix and the global information matrix to obtain an aggregated representation of the local information matrix and an aggregated representation of the global information matrix; An optimal feature learning module that respectively perturbs the aggregated representation of the local information matrix and the aggregated representation of the global information matrix, and performs contrastive learning on the perturbed aggregated representation of the local information matrix and the aggregated representation of the global information matrix to obtain optimal local information features and optimal global information features; A feature optimization module that aggregates the optimal local information features and the optimal global information features to obtain an optimized node representation matrix; A clustering mapping module that performs clustering mapping on the optimized node representation matrix to obtain a preliminary clustering result; A clustering adjustment module that adjusts the preliminary clustering result to obtain a final clustering result.