Graph clustering method and device based on multi-order neighbor information transfer fusion clustering network
Through the multi-order neighbor information transmission and fusion clustering network, combined with random walk and self-attention mechanisms, the graph structure information is optimized, and the problem of poor clustering performance of graph structure data in the existing technology is solved, achieving a more accurate and comprehensive Consistency information indicates that the clustering effect is improved.
Patent Information
- Application Number
- CN202210502357.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-10
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-05-10
AI Technical Summary
The existing deep clustering method based on GCNs lacks the accuracy of the original graph structure when processing graph structure data, and cannot effectively integrate cross-modal dynamic information, resulting in the node representation learning not being comprehensive and accurate enough, affecting clustering performance.
Multi-order neighbor information transmission fusion clustering network is adopted, and the adjacency matrix is optimized through random walks, and a deep fusion clustering network is built, combining node characteristics and topological structure information to perform information fusion, and a self-attention mechanism is introduced to weight topological structure information. Generate soft allocation and target distribution matrix for K-means clustering.
Improve the accuracy and comprehensiveness of clustering. By filtering error relationships, we enhance information transmission between nodes, learn higher-quality node representations, and improve clustering performance.
Smart Images

Figure CN114792113B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data processing technology, and in particular to a graph clustering method and device based on a multi-order neighbor information transmission fusion clustering network. Background Art
[0002] With the continuous development of deep learning technology, deep clustering can make full use of unlabeled data through neural networks to divide data into multiple non-overlapping groups. Since there is no need for time-consuming and labor-intensive data labeling, deep clustering methods have received widespread attention and achieved great success in various applications such as face recognition, social network analysis, and anomaly detection. The two factors of optimizing objectives and feature extraction methods play an important role in improving the performance of deep clustering methods. Specifically, in unsupervised clustering scenarios, without label guidance, it is extremely critical and challenging to design a sophisticated objective function and network architecture so that the network can collect more comprehensive and discriminative information to reveal the intrinsic structure of the data.
[0003] In recent years, graph-structured data has shown great advantages in deep learning. However, due to the non-Euclidean topological graph structure and complex node attributes, most existing deep clustering methods cannot directly analyze such graph-structured data. Therefore, Graph Convolutional Networks (GCNs) have been proposed and have received a lot of attention and development. GCNs are designed to aggregate information from neighboring nodes for node representation learning and show powerful capabilities in node-oriented clustering tasks. For example: Deep Attention Embedded Graph Clustering (DAEGC), which encodes node attributes and topological structures into compact embeddings and reconstructs the adjacent matrix through a self-optimizing embedding method; Following the principle of DAEGC, Adversarial Regularized Graph Autoencoder (ARGA), learns node representation under the guidance of adversarial learning mechanism; Multi-view Graph Representation Learning (MVGRL) uses GCNs-based encoders for cross-view contrastive learning to further improve the quality of node representation, thereby enhancing clustering performance.
[0004] Existing GCNs-based methods still have some limitations. First, most methods rarely consider the accuracy of the original graph structure. In the process of GCNs aggregating neighborhood messages, these unreliable connections may mislead node representation learning and further affect clustering performance. Second, existing methods lack cross-modal dynamic information fusion and processing mechanisms. They directly connect or align the two learning representations from node information and structural information, resulting in insufficient information interaction and merging. Third, the target distribution generation process of existing methods does not make full use of information from two different sources, resulting in incomplete and inaccurate network training, hindered interaction between the two information sources, and unsatisfactory clustering performance. Summary of the invention
[0005] Based on this, it is necessary to provide a graph clustering method and device based on a multi-order neighbor information transmission fusion clustering network to address the above technical problems.
[0006] A graph clustering method based on multi-order neighbor information transmission fusion clustering network, the method comprising:
[0007] Perform random walks on the original graph data samples to obtain the optimized adjacency matrix;
[0008] Constructing a deep fusion clustering network; the deep fusion clustering network includes a first graph encoder and a second graph encoder, and obtaining node feature encoding information of the original graph data sample through the first graph encoder; obtaining global topological structure information of the original graph data sample through the second graph encoder, and using the optimized adjacency matrix as a weight to weight the global topological structure information to obtain structural encoding information of the original graph data sample;
[0009] Fusing the node feature coding information and the structure coding information to obtain fused coding information;
[0010] A soft allocation matrix and a target distribution matrix are constructed for the fused coding information to obtain a soft allocation matrix; the soft allocation matrix is used to represent the sample distribution of the original image data samples; and a target distribution matrix is generated through the soft allocation matrix;
[0011] The deep fusion clustering network is trained according to a preset loss function, and K-means clustering is performed using the node representation corresponding to the trained deep fusion clustering network output graph data.
[0012] In one embodiment, the method further includes: for each node in the original graph data sample, performing k-step random walks starting from the node, and repeating the steps m times to obtain a random walk matrix; the elements in the random walk matrix are determined according to information of whether the node passes through a preset node during the walk;
[0013] According to the random walk matrix, an average walk matrix is calculated; the average walk matrix represents the average number of times a preset node is reached during random walk;
[0014] According to the average walk matrix, an optimized adjacency matrix is obtained.
[0015] In one embodiment, the method further includes: mapping the node features in the original graph data sample in different ways to obtain a query matrix Q and a key matrix K as follows:
[0016]
[0017]
[0018] Among them, l represents the number of network layers, represents the network parameters of layer l, Z (l-1 ) represents the output of the previous network layer, Z (0) is the node feature matrix X, σ is the nonlinear activation function Tanh;
[0019] According to the query matrix Q and the key matrix K, the global structure topology information is obtained as follows:
[0020]
[0021] Among them, d (l) Represents the dimension of the lth hidden layer representation.
[0022] In one embodiment, the method further includes: using the optimized adjacency matrix as a weight to weight the global topological structure information as follows:
[0023]
[0024] Among them, ⊙ represents the Hadamard product, represents the optimized adjacency matrix, G (l) Represents global topology information, Represents weighted global topological structure information;
[0025] According to the key matrix K, the value matrix V is obtained as:
[0026]
[0027] in, represents the network parameters of the lth layer, σ is the nonlinear activation function Tanh;
[0028] A self-attention mechanism is introduced into the weighted global topological structure information and the value matrix V to obtain the structural encoding information:
[0029]
[0030]
[0031] Among them, Z (l) Represents structural encoding information.
[0032] In one embodiment, the method further includes: fusing the node feature coding information and the structure coding information to obtain fused coding information:
[0033] Z I =αZ AE +(1-α)Z TGAE ,
[0034]
[0035] Among them, α is a learnable parameter used to balance the importance of the two representations. represents the fusion coding information, Z AE Represents node feature encoding information, Z TGAE Represents structural encoding information.
[0036] In one embodiment, the method further includes: performing soft allocation matrix calculation on the fused coding information to obtain:
[0037]
[0038] Among them, q ij Indicates that the i-th node in the soft assignment matrix belongs to the j-th pre-computed cluster center u j The probability of , v represents the degrees of freedom of the distribution;
[0039] The target distribution is generated for the fused coding information, and the target allocation matrix is obtained as follows:
[0040]
[0041] Specifically, 0≤p ij ≤1 is the generated target distribution P∈R N An element in ×K represents the probability that the i-th sample belongs to the j-th cluster center.
[0042] In one of the embodiments, the loss function includes: a first reconstruction loss, a second reconstruction loss, and a clustering loss;
[0043] The first reconstruction loss is:
[0044]
[0045] in, To reconstruct the node feature matrix;
[0046] The second reconstruction loss is:
[0047]
[0048] Among them, γ is a predefined hyperparameter, Represents the reconstructed weighted node feature matrix and represents the reconstructed adjacency matrix;
[0049] The clustering loss is:
[0050]
[0051] in, Q', Q" respectively represent the node feature encoding information Z AE and the structure encoding information Z TGAE The corresponding soft assignment matrix.
[0052] A graph clustering device based on a multi-order neighbor information transmission fusion clustering network, the device comprising:
[0053] The preprocessing module is used to perform random walks on the original graph data samples to obtain the optimized adjacency matrix;
[0054] A network construction module, used to construct a deep fusion clustering network; the deep fusion clustering network includes a first graph encoder and a second graph encoder, through which the node feature encoding information of the original graph data sample is obtained; through the second graph encoder, the global topological structure information of the original graph data sample is obtained, and the optimized adjacency matrix is used as a weight to weight the global topological structure information to obtain the structural encoding information of the original graph data sample;
[0055] A fusion module, used for fusing the node feature coding information and the structure coding information to obtain fused coding information;
[0056] A distribution matrix construction module is used to construct a soft allocation matrix and a target distribution matrix for the fusion coding information to obtain a soft allocation matrix; the soft allocation matrix is used to represent the sample distribution of the original image data samples; and then the target distribution matrix is generated through the soft allocation matrix;
[0057] The clustering module is used to train the deep fusion clustering network according to a preset loss function, and perform K-means clustering using the node representation corresponding to the trained deep fusion clustering network output graph data.
[0058] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:
[0059] Perform random walks on the original graph data samples to obtain the optimized adjacency matrix;
[0060] Constructing a deep fusion clustering network; the deep fusion clustering network includes a first graph encoder and a second graph encoder, and obtaining node feature encoding information of the original graph data sample through the first graph encoder; obtaining global topological structure information of the original graph data sample through the second graph encoder, and using the optimized adjacency matrix as a weight to weight the global topological structure information to obtain structural encoding information of the original graph data sample;
[0061] Fusing the node feature coding information and the structure coding information to obtain fused coding information;
[0062] A soft allocation matrix and a target distribution matrix are constructed for the fused coding information to obtain a soft allocation matrix; the soft allocation matrix is used to represent the sample distribution of the original image data samples; and a target distribution matrix is generated through the soft allocation matrix;
[0063] The deep fusion clustering network is trained according to a preset loss function, and K-means clustering is performed using the node representation corresponding to the trained deep fusion clustering network output graph data.
[0064] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:
[0065] Perform random walks on the original graph data samples to obtain the optimized adjacency matrix;
[0066] Constructing a deep fusion clustering network; the deep fusion clustering network includes a first graph encoder and a second graph encoder, and obtaining node feature encoding information of the original graph data sample through the first graph encoder; obtaining global topological structure information of the original graph data sample through the second graph encoder, and using the optimized adjacency matrix as a weight to weight the global topological structure information to obtain structural encoding information of the original graph data sample;
[0067] Fusing the node feature coding information and the structure coding information to obtain fused coding information;
[0068] A soft allocation matrix and a target distribution matrix are constructed for the fused coding information to obtain a soft allocation matrix; the soft allocation matrix is used to represent the sample distribution of the original image data samples; and a target distribution matrix is generated through the soft allocation matrix;
[0069] The deep fusion clustering network is trained according to a preset loss function, and K-means clustering is performed using the node representation corresponding to the trained deep fusion clustering network output graph data.
[0070] The above-mentioned graph clustering method, device, computer equipment and storage medium based on multi-order neighbor information transmission fusion clustering network, on the one hand, can filter out the wrong relationships in the original adjacency or reduce the weight of these wrong relationships through random walk preprocessing, while adding some reliable relationships, thereby improving the accuracy of clustering, on the other hand, establish a global topological relationship through node feature similarity, and use the newly created adjacency matrix as the topological weight, so that the topological relationship between nodes simultaneously incorporates node feature information and structural information. GCNs with the addition of self-attention mechanism can transfer information between nodes at a longer distance in a single-layer network, learn higher quality node representations, and finally further improve the accuracy of clustering through information fusion. BRIEF DESCRIPTION OF THE DRAWINGS
[0071] Figure 1 A schematic diagram of a flow chart of a graph clustering method based on multi-order neighbor information transmission and fusion clustering network in one embodiment;
[0072] Figure 2 It is a structural block diagram of a graph clustering device based on multi-order neighbor information transmission fusion clustering network in one embodiment;
[0073] Figure 3 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION
[0074] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0075] In one embodiment, Figure 1 As shown, a graph clustering method based on multi-order neighbor information transmission fusion clustering network is provided, comprising the following steps:
[0076] Step 102: Perform a random walk on the original graph data samples to obtain an optimized adjacency matrix.
[0077] Random walk refers to randomly walking k steps from node i, and walking multiple times. By counting the average number of times each node is passed, the relationship between the local neighbor nodes of node i is reflected. By calculating the similarity of the local neighbor information of each node, a new adjacency relationship is obtained, thereby filtering out the erroneous relationships in the original adjacency matrix.
[0078] Step 104, construct a deep fusion clustering network.
[0079] The deep fusion clustering network includes a first graph encoder and a second graph encoder. Specifically, the first graph encoder is an autoencoder AE, and the second graph encoder is a transformer-based graph autoencoder TGAE.
[0080] Specifically, the self-attention transformer network (Transformer) is composed of a self-attention mechanism (Self-attention) and a feedforward neural network. It first maps the data features to three matrices Q (Query), K (Key), and V (Value), and then obtains the global topological structure by matrix multiplication of Q and K, as shown in the following formula, where d k is the dimension of K.
[0081]
[0082] Through this formula, the latent space data representation can be obtained through the self-attention mechanism.
[0083] The node feature encoding information of the original graph data sample is obtained through the first graph encoder; the global topological structure information of the original graph data sample is obtained through the second graph encoder, and the optimized adjacency matrix is used as the weight to weight the global topological structure information to obtain the structural encoding information of the original graph data sample.
[0084] Step 106, fusing the node feature coding information and the structure coding information to obtain fused coding information.
[0085] Step 108, constructing a soft allocation matrix and a target distribution matrix for the fused coding information to obtain a soft allocation matrix;
[0086] The soft allocation matrix is used to represent the sample distribution of the original graph data samples; and the target distribution matrix is generated through the soft allocation matrix.
[0087] Step 110, train the deep fusion clustering network according to a preset loss function, and perform K-means clustering using the node representation matrix corresponding to the trained deep fusion clustering network output graph data.
[0088] In the above graph clustering method based on multi-order neighbor information transmission fusion clustering network, on the one hand, through random walk preprocessing, the erroneous relationships in the original adjacency can be filtered out or the weights of these erroneous relationships can be reduced, while adding some reliable relationships. On the other hand, the global topological relationship is established through the node feature similarity, and the newly created adjacency matrix is used as the topological weight, so that the topological relationship between nodes is integrated with the node feature information and structural information at the same time. GCNs with the addition of self-attention mechanism can transfer information between nodes at a longer distance in a single-layer network, learn higher-quality node representations, and finally further improve the accuracy of clustering through information fusion.
[0089] In one of the embodiments, for each node in the original graph data sample, a k-step random walk is performed starting from the node and repeated m times to obtain a random walk matrix; the elements in the random walk matrix are determined based on information about whether a preset node is passed during the walk; based on the random walk matrix, an average walk matrix is calculated; the average walk matrix represents the average number of times the preset nodes are reached during the random walk; based on the average walk matrix, an optimized adjacency matrix is obtained.
[0090] Specifically, given an undirected graph is a set of points, is the edge set, M, A is the original adjacency matrix, A ij =1 means there is an edge between node i and node j, otherwise A ij = 0. By performing k steps of random walk from each node and repeating m times, m random walk matrices F = F1,…,F m ,F a ∈R N×N represents the a-th random walk matrix, It means that in the ath random walk, it passes through node j from node i, otherwise it is 0. Then we get the average walk matrix f i,j represents the average number of times from node i to node j after m random walks, f i It represents the k-order neighbor information of node i. Finally, through Get the new adjacency relationship between nodes i and j.
[0091] In one embodiment, the node features in the original graph data sample are mapped differently to obtain the query matrix Q and the key matrix K as follows:
[0092]
[0093]
[0094] Among them, l represents the number of network layers, represents the network parameters of layer l, Z (l-1) Represents the output of the previous network layer, Z (0) is the node feature matrix X, σ is the nonlinear activation function Tanh;
[0095] According to the query matrix Q and the bond matrix K, the global structural topology information is obtained as follows:
[0096]
[0097] Among them, d (l) Represents the dimension of the lth hidden layer representation.
[0098] In another embodiment, the optimized adjacency matrix is used as the weight to weight the global topological structure information as follows:
[0099]
[0100] Among them, ⊙ represents the Hadamard product, represents the optimized adjacency matrix, G (l) Represents global topology information, Represents weighted global topological structure information;
[0101] According to the key matrix K, the value matrix V is obtained as:
[0102]
[0103] in, represents the network parameters of the lth layer, σ is the nonlinear activation function Tanh;
[0104] The self-attention mechanism is introduced into the weighted global topological structure information and the value matrix V, and the structural encoding information is obtained as follows:
[0105]
[0106]
[0107] Among them, Z (l) Represents structural encoding information.
[0108] In one embodiment, the node feature coding information and the structure coding information are fused to obtain the fused coding information:
[0109] Z I =αZ AE +(1-α)Z TGAE ,
[0110]
[0111] Among them, α is a learnable parameter used to balance the importance of the two representations. represents the fusion coding information, Z AE Represents node feature encoding information, Z TGAE Represents structural encoding information.
[0112] Specifically, fusion can be achieved through the Structure and Attribute Information Fusion (SAIF) module.
[0113] In one embodiment, a soft allocation matrix is calculated for the fused coding information:
[0114]
[0115] Among them, q ij Indicates that the i-th node in the soft assignment matrix belongs to the j-th pre-computed cluster center u j The probability of , v represents the degrees of freedom of the distribution.
[0116] Specifically, the first step is to use Student's t distribution to calculate the i-th sample in the fusion embedding space and the jth precomputed cluster center u j The similarity between them, v is the degrees of freedom of the Student's t distribution, q ij represents the probability of assigning the i-th node to the j-th center (i.e., soft assignment). Therefore, the query matrix Q∈R N×K reflects the distribution of all samples, and K is the number of cluster centers. In the second step, in order to increase the confidence of cluster assignment, we introduce the following formula to drive all samples closer to the cluster center:
[0117]
[0118] Specifically, 0≤p ij ≤1 is the generated target distribution P∈R N×K An element in represents the probability that the i-th sample belongs to the j-th cluster center.
[0119] and Generate the query matrix Q by Z TGAE and Z AE Two query matrices Q’, Q’ can be generated. In order to train the network in a unified framework and improve the expressiveness of each component, we design a triplet clustering loss by adopting the following form of KL-divergence:
[0120]
[0121] In this formula, Z AE , Z TGAE and fusion representation The sum of the soft assignment distributions of is simultaneously aligned with the target distribution. Since the target distribution is generated without human guidance, we name the loss function as triplet clustering loss and the corresponding training mechanism as triplet self-supervised strategy.
[0122] In one embodiment, the loss function includes three parts: reconstruction loss of AE, reconstruction loss of TGAE and KL loss, and λ is a predefined hyperparameter:
[0123] L=L AE +L TGAE +λL KL .
[0124] L AE It is the reconstruction loss of AE, taking the node feature matrix X as input and obtaining the hidden layer representation Z through the encoder. AE , then expressed as a fusion As the input of the decoder, the reconstructed node feature matrix is obtained
[0125]
[0126] L TGAE is the reconstruction loss of TGAE, with the node feature matrix X and the adjacency matrix constructed by random walk As input, the encoder obtains the hidden layer representation Z TGAE , then expressed as a fusion As the input of the decoder, the reconstructed weighted node feature matrix is obtained and the reconstructed adjacency matrix
[0127]
[0128] Here, γ is a predefined hyperparameter.
[0129] L KL As explained above, the clustering loss is:
[0130]
[0131] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.
[0132] In one embodiment, Figure 2 As shown, a graph clustering device based on multi-order neighbor information transmission fusion clustering network is provided, including: a preprocessing module 202, a network construction module 204, a fusion module 206, a distribution matrix construction module 208 and a clustering module 210, wherein:
[0133] The preprocessing module is used to perform random walks on the original graph data samples to obtain the optimized adjacency matrix;
[0134] A network construction module, used to construct a deep fusion clustering network; the deep fusion clustering network includes a first graph encoder and a second graph encoder, through which the node feature encoding information of the original graph data sample is obtained; through the second graph encoder, the global topological structure information of the original graph data sample is obtained, and the optimized adjacency matrix is used as a weight to weight the global topological structure information to obtain the structural encoding information of the original graph data sample;
[0135] A fusion module, used for fusing the node feature coding information and the structure coding information to obtain fused coding information;
[0136] A distribution matrix construction module is used to construct a soft allocation matrix and a target distribution matrix for the fusion coding information to obtain a soft allocation matrix; the soft allocation matrix is used to represent the sample distribution of the original image data samples; and then the target distribution matrix is generated through the soft allocation matrix;
[0137] The clustering module is used to train the deep fusion clustering network according to a preset loss function, and perform K-means clustering using the node representation corresponding to the trained deep fusion clustering network output graph data.
[0138] For the specific definition of the graph clustering device based on the multi-order neighbor information transmission fusion clustering network, please refer to the definition of the graph clustering method based on the multi-order neighbor information transmission fusion clustering network mentioned above, which will not be repeated here. Each module in the above-mentioned graph clustering device based on the multi-order neighbor information transmission fusion clustering network can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
[0139] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a graph clustering method based on a multi-order neighbor information transmission fusion clustering network is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the computer device shell, or an external keyboard, touchpad or mouse, etc.
[0140] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0141] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method in the above embodiment when executing the computer program.
[0142] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method in the above embodiment are implemented.
[0143] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).
[0144] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0145] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.
Claims
1. A graph clustering method based on multi-order neighbor information transmission fusion clustering network, characterized in that: The method comprises: Performing random walks on the original graph data samples to obtain an optimized adjacency matrix; constructing a deep fusion clustering network; the deep fusion clustering network includes a first graph encoder and a second graph encoder, and obtaining node feature encoding information of the original graph data samples through the first graph encoder; obtaining global topological structure information of the original graph data samples through the second graph encoder, and using the optimized adjacency matrix as a weight to weight the global topological structure information to obtain structural encoding information of the original graph data samples; The node feature coding information and the structure coding information are fused to obtain fused coding information, which is the cluster representation and is also the input of two decoders; Constructing a soft allocation matrix and a target distribution matrix for the fused coding information; the soft allocation matrix is used to represent the sample distribution of the original image data samples, and the target distribution matrix is generated through the soft allocation matrix; According to a preset loss function, the deep fusion clustering network is trained, and the node representation corresponding to the output graph data of the trained deep fusion clustering network is used to perform K-means clustering; Acquiring global topological structure information of the original graph data sample through the second graph encoder includes: The node features in the original graph data sample are mapped differently to obtain the query matrix Q and the key matrix K as follows: Among them, l represents the number of network layers, represents the network parameters of layer l, Z (l-1) Represents the output of the previous network layer, Z (0) is the node feature matrix X, σ is the nonlinear activation function Tanh; According to the query matrix Q and the key matrix K, the global structure topology information is obtained as follows: Among them, d (l) Indicates the dimension of the lth hidden layer representation; The optimized adjacency matrix is used as a weight to weight the global topological structure information to obtain the structural coding information of the original graph data sample, including: The optimized adjacency matrix is used as the weight to weight the global topological structure information as follows: Among them, ⊙ represents the Hadamard product, represents the optimized adjacency matrix, G (l) Represents global topology information, Represents weighted global topological structure information; According to the key matrix K, the value matrix V is obtained as: in, represents the network parameters of the lth layer, σ is the nonlinear activation function Tanh; A self-attention mechanism is introduced into the weighted global topological structure information and the value matrix V to obtain the structural encoding information: Among them, Z (l) Representation structure encoding information; Calculating a soft allocation matrix for the fused coding information; the soft allocation matrix is used to represent the sample distribution of the original image data samples, and generating a target distribution matrix through the soft distribution matrix, including: The soft allocation matrix is calculated for the fused coding information to obtain the soft allocation matrix: Among them, q ij Indicates that the i-th node in the soft assignment matrix belongs to the j-th pre-computed cluster center u j The probability of , v represents the degrees of freedom of the distribution; The target distribution is generated for the fused coding information, and the target allocation matrix is obtained as follows: Where 0≤p ij ≤1 is the generated target distribution P∈R N×K An element in represents the probability that the i-th sample belongs to the j-th cluster center.
2. The method according to claim 1, characterized in that The random walk of the original graph data samples to obtain an optimized adjacency matrix includes: For each node in the original graph data sample, a k-step random walk is performed starting from the node, and repeated m times to obtain a random walk matrix; the elements in the random walk matrix are determined according to whether the preset node is passed during the walk; According to the random walk matrix, an average walk matrix is calculated; the average walk matrix represents the average number of times a preset node is reached during random walk; According to the average walk matrix, an optimized adjacency matrix is obtained.
3. The method according to claim 1, characterized in that The node feature coding information and the structure coding information are fused to obtain fused coding information, including: The node feature coding information and the structure coding information are fused to obtain the fused coding information: WITH I =αZ AE +(1-α)Z TGAE , Among them, α is a learnable parameter used to balance the importance of the two representations. represents the fusion coding information, Z AE Represents node feature encoding information, Z TGAE Represents structural encoding information.
4. The method according to any one of claims 1 to 3, characterized in that: The loss function includes: a first reconstruction loss, a second reconstruction loss and a clustering loss; The first reconstruction loss is: in, To reconstruct the node feature matrix; The second reconstruction loss is: Among them, γ is a predefined hyperparameter, Represents the reconstructed weighted node feature matrix and represents the reconstructed adjacency matrix; The clustering loss is: in, Q', Q" respectively represent the node feature encoding information Z AE and the structure encoding information Z TGAE The corresponding soft assignment matrix.
5. A graph clustering device based on multi-order neighbor information transmission fusion clustering network, characterized in that: The device comprises: The preprocessing module is used to perform random walks on the original graph structure to obtain an optimized adjacency matrix; A network construction module, used to construct a deep fusion clustering network; the deep fusion clustering network includes a first graph encoder and a second graph encoder, through which the node feature encoding information of the original graph data sample is obtained; through the second graph encoder, the global topological structure information of the original graph data sample is obtained, and the optimized adjacency matrix is used as a weight to weight the global topological structure information to obtain the structural encoding information of the original graph data sample; A fusion module, used for fusing the node feature coding information and the structure coding information to obtain fused coding information; A soft allocation matrix and target distribution matrix construction module is used to construct a soft allocation matrix and a target distribution matrix for the fusion coding information; the soft allocation matrix is used to represent the sample distribution of the original image data samples, and the target distribution matrix is generated through the soft distribution matrix; A clustering module, used to train the deep fusion clustering network according to a preset loss function, and perform K-means clustering using the node representation corresponding to the trained deep fusion clustering network output graph data; Acquiring global topological structure information of the original graph data sample through the second graph encoder includes: The node features in the original graph data sample are mapped differently to obtain the query matrix Q and the key matrix K as follows: Among them, l represents the number of network layers, represents the network parameters of layer l, Z (l-1) Represents the output of the previous network layer, Z (0) is the node feature matrix X, σ is the nonlinear activation function Tanh; According to the query matrix Q and the key matrix K, the global structure topology information is obtained as follows: Among them, d (l) Indicates the dimension of the lth hidden layer representation; The optimized adjacency matrix is used as a weight to weight the global topological structure information to obtain the structural coding information of the original graph data sample, including: The optimized adjacency matrix is used as the weight to weight the global topological structure information as follows: Among them, ⊙ represents the Hadamard product, represents the optimized adjacency matrix, G (l) Represents global topology information, Represents weighted global topological structure information; According to the key matrix K, the value matrix V is obtained as: in, represents the network parameters of the lth layer, σ is the nonlinear activation function Tanh; A self-attention mechanism is introduced into the weighted global topological structure information and the value matrix V to obtain the structural encoding information: Among them, Z (l) Representation structure encoding information; Calculating a soft allocation matrix for the fused coding information; the soft allocation matrix is used to represent the sample distribution of the original image data samples, and generating a target distribution matrix through the soft distribution matrix, including: The soft allocation matrix is calculated for the fused coding information to obtain the soft allocation matrix: Among them, q ij Indicates that the i-th node in the soft assignment matrix belongs to the j-th pre-computed cluster center u j The probability of , v represents the degrees of freedom of the distribution; The target distribution is generated for the fused coding information, and the target allocation matrix is obtained as follows: Where 0≤p ij ≤1 is the generated target distribution P∈R N×K An element in represents the probability that the i-th sample belongs to the j-th cluster center.
6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 4 are implemented.
7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.
