A deep graph clustering method based on dual correlation reduction
By performing data augmentation and bicorrelation reduction processing on graph data, the representation collapse problem in graph convolution neural network clustering algorithm is solved, the discriminant ability and clustering performance of node representation are improved, and memory space is saved.
Patent Information
- Application Number
- CN202210332349.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-31
- Publication Date
- 2024-12-10
- Estimated Expiration
- 2042-03-31
AI Technical Summary
Traditional clustering algorithms based on graph convolution neural networks have a problem of representation collapse, which leads to insufficient discriminant ability of node representation, limiting clustering performance, especially complex negative sample generation operations occupy memory space and affecting the GPU running speed.
Using the graph depth clustering method based on bicorrelation reduction, data augmentation is enhanced by obtaining the original target relationship graph, classify the adjacency matrix and diffusion matrix, combine Gaussian noise to generate the augmented graph, and input it to pre-trained bicorrelation Reduce the correlation at the sample level and feature level of the network to reduce the correlation at the sample level, avoid representation collapse, and save memory space.
It improves the discriminant ability of node representation, avoids representation collapse, improves clustering performance, and saves memory space during graph clustering, reducing the complexity of negative sample generation.
Smart Images
Figure CN114861757B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of deep learning technology, and in particular to a graph deep clustering method based on dual correlation reduction. Background Art
[0002] Deep graph clustering is to train a deep neural network to learn the embedding representation of nodes to divide nodes into different clusters without manual annotation. Due to its powerful graph information capture ability, graph convolutional networks (GCNs) have recently achieved promising performance in many graph clustering applications such as social networks and recommendation systems. Thanks to its powerful graph information extraction ability, GNNs have achieved promising results in deep graph clustering, but the problem of over-smoothing remains unsolved. Since then, Structural Deep Clustering Network (SDCN) / Structural Deep Clustering Network with Q Distribution (SDCN_Q) and Deep Fusion Clustering Network (DFCN) have proposed jointly training autoencoders (AE) and GAEs to alleviate the over-smoothing problem through information transfer operations and structural attribute fusion modules, respectively.
[0003] However, traditional clustering algorithms based on graph convolutional neural networks usually have the problem of representation collapse, that is, they tend to map nodes of different categories to similar representations during sample embedding. For example, GAE and MVGRL (Multi-View Graph Representation Learning), GAE is a classic graph convolutional network, and MVGRL is an algorithm enhanced by contrastive learning strategy. By introducing a positive and negative sample pair recognition mechanism, the problem of representation collapse can be alleviated to a certain extent. However, the complex negative sample generation operation will occupy memory space and affect the GPU running speed. In this case, the lack of discriminative ability of node representation will lead to limited clustering performance, which limits the performance of existing GCN-based clustering algorithms. Summary of the invention
[0004] Based on this, it is necessary to provide a graph deep clustering method, device, computer equipment and storage medium based on double correlation reduction to address the above technical problems.
[0005] A graph deep clustering method based on dual correlation reduction, the method comprising:
[0006] Obtain the original target relationship graph to be clustered;
[0007] Based on feature similarity and graph diffusion, data enhancement is performed on the original target relationship graph to obtain a classification graph adjacency matrix and a diffusion matrix, a first augmented graph is obtained according to the classification graph adjacency matrix and Gaussian noise, and a second augmented graph is obtained according to the diffusion matrix and Gaussian noise;
[0008] Inputting the feature graphs of the first augmented graph and the second augmented graph into a pre-trained dual correlation reduction network to perform sample-level correlation reduction and feature-level correlation reduction, respectively, to obtain a dual correlation reduction feature graph;
[0009] The dual correlation reduction feature graph is clustered to obtain a target relationship clustering graph.
[0010] In one embodiment, the method further includes: calculating the cosine similarity between the nodes of the original target relationship graph to obtain an adjacency matrix based on similarity; and classifying the adjacency matrix based on similarity using a KNN algorithm to obtain a classified graph adjacency matrix:
[0011]
[0012] in, Represents the normalization function, represents the K-nearest neighbor algorithm, The feature matrix representing the original target relationship graph, Represents the transpose operation of a matrix.
[0013] In one embodiment, the method further includes: performing generalized graph diffusion on the adjacency matrix of the original target relationship graph to obtain a diffusion matrix:
[0014]
[0015] in, is the transmission probability, represents the identity matrix, , is the adjacency matrix of the original target relationship graph, is the normalized adjacency matrix, The feature matrix representing the original target relationship graph, Represents the matrix inversion operation.
[0016] In one of the embodiments, it also includes: obtaining a cross-view sample correlation matrix based on the feature graphs of the first augmented graph and the second augmented graph; constructing a sample-level mean square error loss function based on the cross-view sample correlation matrix and the clustering optimization adjacency matrix; when the sample-level mean square error loss function is less than a preset threshold, obtaining a sample-level correlation reduction feature graph.
[0017] In one embodiment, the method further includes: obtaining cluster-level feature maps according to the feature maps of the first augmented map and the second augmented map respectively; obtaining a cross-view feature correlation matrix according to the cluster-level feature maps of the first augmented map and the second augmented map; and obtaining a cross-view feature correlation matrix according to the cross-view feature correlation matrix and the unit matrix. , construct a feature layer mean square error loss function; when the feature layer mean square error loss function is less than a preset threshold, obtain a feature layer correlation reduction feature map.
[0018] In one embodiment, the method further includes: constructing a sample-level mean square error loss function according to the cross-view sample correlation matrix and the cluster optimization adjacency matrix:
[0019]
[0020] in, represents the cross-view sample correlation matrix, represents the clustering optimized adjacency matrix, , If is equal to 1, If is equal to 0, yes and The cosine similarity between and Denote the i-th node embedding of the first augmented view and the j-th node embedding of the second augmented view.
[0021] In one embodiment, the method further includes: according to the cross-view feature correlation matrix and the identity matrix , construct the feature layer mean square error loss function as:
[0022]
[0023] in, is the dimension of the learned latent embedding, , represents the cross-view feature correlation matrix, represents the cosine similarity between the i-th dimension feature of the first augmented view and the j-th dimension feature of the second augmented view.
[0024] In one of the embodiments, it also includes: obtaining a dual correlation reduction feature map according to the sample-level correlation reduction feature map and the feature-level correlation reduction feature map.
[0025] A graph deep clustering device based on dual correlation reduction, the device comprising:
[0026] A graph acquisition module, used to obtain the original target relationship graph to be clustered;
[0027] A data enhancement module, used to perform data enhancement on the original target relationship graph based on feature similarity and graph diffusion, respectively, to obtain a classification graph adjacency matrix and a diffusion matrix, to obtain a first augmented graph according to the classification graph adjacency matrix and Gaussian noise, and to obtain a second augmented graph according to the diffusion matrix and Gaussian noise;
[0028] A dual correlation reduction module, used for inputting the feature graphs of the first augmented graph and the second augmented graph into a pre-trained dual correlation reduction network to perform sample-level correlation reduction and feature-level correlation reduction, respectively, to obtain a dual correlation reduction feature graph;
[0029] The graph generation module is used to perform graph clustering on the dual correlation reduction feature graph to obtain a target relationship clustering graph.
[0030] 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:
[0031] Obtain the original target relationship graph to be clustered;
[0032] Based on feature similarity and graph diffusion, data enhancement is performed on the original target relationship graph to obtain a classification graph adjacency matrix and a diffusion matrix, a first augmented graph is obtained according to the classification graph adjacency matrix and Gaussian noise, and a second augmented graph is obtained according to the diffusion matrix and Gaussian noise;
[0033] Inputting the feature graphs of the first augmented graph and the second augmented graph into a pre-trained dual correlation reduction network to perform sample-level correlation reduction and feature-level correlation reduction, respectively, to obtain a dual correlation reduction feature graph;
[0034] The dual correlation reduction feature graph is clustered to obtain a target relationship clustering graph.
[0035] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:
[0036] Obtain the original target relationship graph to be clustered;
[0037] Based on feature similarity and graph diffusion, data enhancement is performed on the original target relationship graph to obtain a classification graph adjacency matrix and a diffusion matrix, a first augmented graph is obtained according to the classification graph adjacency matrix and Gaussian noise, and a second augmented graph is obtained according to the diffusion matrix and Gaussian noise;
[0038] Inputting the feature graphs of the first augmented graph and the second augmented graph into a pre-trained dual correlation reduction network to perform sample-level correlation reduction and feature-level correlation reduction, respectively, to obtain a dual correlation reduction feature graph;
[0039] The dual correlation reduction feature graph is clustered to obtain a target relationship clustering graph.
[0040] The above-mentioned graph deep clustering method, apparatus, computer equipment and storage medium based on dual correlation reduction can perform data enhancement on the input original target relationship graph based on feature similarity and graph diffusion through a dual correlation reduction network, so as to perturb the graph data so that the model can learn richer information and improve the discriminative ability of node representation. Based on the augmented graph obtained after data enhancement, the correlation at the sample level and feature level is reduced respectively, so that the redundant information of learning embedding can be filtered and the network can be guided to learn more discriminative embedding, thereby avoiding representation collapse. The obtained dual correlation reduction feature graph is subjected to graph clustering to obtain a clustering graph. In the embodiment of the present invention, there is no negative sample generation, which can save memory space in the graph clustering process and improve clustering performance by avoiding representation collapse. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 It is a schematic diagram of a flow chart of a graph deep clustering method based on dual correlation reduction in one embodiment;
[0042] Figure 2 It is a schematic flow chart of a step of reducing the correlation of a first augmented image and a second augmented image at a sample level in one embodiment;
[0043] Figure 3 A schematic diagram of a flow chart of a step of reducing the correlation of a first augmented image and a second augmented image at a feature level in one embodiment;
[0044] Figure 4 It is a schematic diagram of the overall process of a graph deep clustering method based on dual correlation reduction in a specific embodiment;
[0045] Figure 5 A schematic diagram of a DCRM process in one embodiment;
[0046] Figure 6 A schematic diagram of GPU memory cost ablation of a graph deep clustering method based on dual correlation reduction on ACM, CITE and DBLP datasets in one embodiment;
[0047] Figure 7 is a structural block diagram of a graph deep clustering device based on dual correlation reduction in one embodiment;
[0048] Figure 8 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION
[0049] 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.
[0050] In one embodiment, Figure 1 As shown, a graph deep clustering method based on double correlation reduction is provided, comprising the following steps:
[0051] Step 102: Obtain the original target relationship graph to be clustered.
[0052] A graph is a structured data representation based on vertices and edges. A target relationship graph refers to a data model formed by the relationship between targets in a network. A network can be a social network or a product network. Targets in social networks can be users or customers. Clustering in social networks can help find community groups with the same interests. Targets in product networks can be products, users, and advertisers. Clustering users in a recommendation system can re-summarize labels for each user category based on the clustering results to obtain a new user classification system, which can better push products. Not only is this classification system clearer and more reasonable than before, it is also more capable of handling large volumes of users and corresponding user data in the product ecosystem, and can also help advertising platforms and advertisers better understand user groups.
[0053] Step 104, based on feature similarity and graph diffusion, data enhancement is performed on the original target relationship graph to obtain a classification graph adjacency matrix and a diffusion matrix, a first augmented graph is obtained according to the classification graph adjacency matrix and Gaussian noise, and a second augmented graph is obtained according to the diffusion matrix and Gaussian noise.
[0054] Data augmentation of the original target relationship graph can improve the discriminative ability of node representation. The method of the present invention considers two types of augmentation, namely feature perturbation and structure construction, which can be expressed as a data augmentation module (GraphAugmentation Module, GAM). This module perturbs the graph data. The feature perturbation is generated by Gaussian noise. Specifically, a random noise matrix is generated. , which is a Gaussian distribution with a mean of 1 and a standard deviation of 0.1 , calculate the perturbation attribute matrix , the formula is as follows:
[0055]
[0056] in, is the feature matrix of the original target relationship graph, It stands for Hadamard product. Perturbing the graph data can enable the model to learn richer information. Based on feature similarity, the original target relationship graph is enhanced. The graph structure can be constructed through the relationship between the nodes of the original target relationship graph to obtain the classification graph adjacency matrix. The classification graph adjacency matrix is obtained through the classification algorithm. Based on graph diffusion, the original target relationship graph is enhanced to obtain the diffusion matrix, which can provide local and global information of the original target relationship graph. Based on feature perturbation and structural construction, the first augmented graph and the second augmented graph are obtained, which can guide the dual correlation reduction network to learn more discriminative embeddings from the two generated augmented graphs.
[0057] Step 106: Input the feature graphs of the first augmented graph and the second augmented graph into a pre-trained dual correlation reduction network to perform sample-level correlation reduction and feature-level correlation reduction, respectively, to obtain a dual correlation reduction feature graph.
[0058] The sample-level correlation reduction can guide the dual correlation reduction network to learn more discriminative potential features about the sample in the first augmented view and the second augmented view, thereby alleviating the folded representation. The feature-level correlation reduction can reduce the redundant information in the learned representation, thereby improving the discriminative ability of the feature. The sample-level correlation reduction (SCRM) and the feature-level correlation reduction (FCRM) can be combined to represent the dual correlation reduction module (DCRM). The dual correlation reduction strategy frees the method of the present invention from complex negative sample generation operations, thus saving more space and being more flexible to the training batch size. The dual correlation reduction network includes a data enhancement structure and a dual correlation reduction structure, and is a twin network that can solve the representation collapse problem in the field of deep graph clustering.
[0059] Step 108, performing graph clustering on the dual correlation reduction feature graph to obtain a target relationship clustering graph.
[0060] The dual correlation reduction feature graph is clustered through the K-means algorithm and clustering guidance to obtain the target relationship clustering graph. The clustering algorithms used also include Graph Community Detection and mean shift clustering. Clustering guidance uses clustering results with high confidence to guide clustering results with low confidence, using KL divergence as the loss function, which can make the clustering results more accurate.
[0061] In the above-mentioned graph deep clustering method based on dual correlation reduction, by using a dual correlation reduction network, data enhancement is performed on the input original target relationship graph based on feature similarity and graph diffusion, so that the graph data can be disturbed, so that the model can learn richer information and improve the discriminative ability of node representation. Based on the augmented graph obtained after data enhancement, the correlation at the sample level and the feature level is reduced respectively, so that the redundant information of the learning embedding can be filtered out, and the dual correlation reduction network is guided to learn a more discriminative embedding, thereby avoiding representation collapse. The obtained dual correlation reduction feature graph is subjected to graph clustering to obtain a clustering graph. In the embodiment of the present invention, there is no negative sample generation, which can save memory space in the graph clustering process and improve clustering performance by avoiding representation collapse.
[0062] In one embodiment, based on feature similarity, data enhancement is performed on the original target relationship graph to obtain a classification graph adjacency matrix, including: calculating the cosine similarity between the nodes of the original target relationship graph to obtain an adjacency matrix based on similarity; and classifying the adjacency matrix based on similarity using a KNN algorithm to obtain a classification graph adjacency matrix:
[0063]
[0064] in, Represents the normalization function, represents the K-nearest neighbor algorithm, The feature matrix representing the original target relationship graph, Represents the transpose operation of a matrix.
[0065] In this embodiment, the KNN graph can be obtained by calculating the cosine similarity between the nodes in the original target relationship graph and constructing the structure through the K-nearest neighbor algorithm. Specifically, the number of nearest neighbors in the KNN algorithm is Set to 5.
[0066] In one embodiment, based on graph diffusion, data enhancement is performed on the original target relationship graph to obtain a diffusion matrix, which includes: performing generalized graph diffusion on the adjacency matrix of the original target relationship graph to obtain a diffusion matrix:
[0067]
[0068] in, is the transmission probability, represents the identity matrix, , is the adjacency matrix of the original target relationship graph, is the normalized adjacency matrix, The feature matrix representing the original target relationship graph, Represents the matrix inversion operation.
[0069] In this embodiment, is the transmission probability, which is taken as 0.2. Through the generalized graph diffusion in MVGRL, the adjacency matrix is converted into a diffusion matrix, and the two matrices are regarded as two matrices of the same graph structure. Contract View,Since the adjacency matrix and diffusion matrix provide local and global views of the graph structure, respectively, the consistency between the representations learned from these two views is maximized, allowing the model to encode rich local and global information simultaneously.
[0070] In one embodiment, if Figure 2 As shown, a schematic flowchart of a step of reducing the correlation of a first augmented image and a second augmented image at a sample level is provided, and the specific steps are as follows:
[0071] Step 202: Obtain a cross-view sample correlation matrix according to the feature maps of the first augmented image and the second augmented image.
[0072] Step 204: construct a sample-level mean square error loss function based on the cross-view sample correlation matrix and the cluster-optimized adjacency matrix.
[0073] Step 206, when the sample-level mean square error loss function is less than a preset threshold, a sample-level correlation reduction feature map is obtained.
[0074] In this embodiment, when the sample layer mean square error loss function is less than a preset threshold, the cross-view sample correlation matrix approaches the clustering optimized adjacency matrix, and the sample-level correlation reduction mechanism proposed in the present invention guides the dual correlation reduction network to learn more discriminative potential features about the samples in the two enhanced views, thereby alleviating the collapsed representation.
[0075] Specifically, the first augmented graph and the second augmented graph Node embeddings encoded as two views and , calculate the cross-view sample correlation matrix:
[0076]
[0077] The elements yes and The cosine similarity between and Denote the i-th node embedding of the first augmented view and the j-th node embedding of the second augmented view.
[0078] In one embodiment, constructing a sample-level mean square error loss function according to the cross-view sample correlation matrix and the cluster-optimized adjacency matrix includes: constructing a sample-level mean square error loss function according to the cross-view sample correlation matrix and the cluster-optimized adjacency matrix as:
[0079]
[0080] in, represents the cross-view sample correlation matrix, represents the clustering optimized adjacency matrix, , If is equal to 1, If is equal to 0, yes and The cosine similarity between and Denote the i-th node embedding of the first augmented view and the j-th node embedding of the second augmented view.
[0081] like Figure 4 As shown, the adjacency matrix for clustering optimization is designed The steps are:
[0082] The first step is to consider the homogeneity of neighbors in the graph and use initialization , is an adjacency matrix with self-loops, i.e. Figure 4 Step 1:Init with self-looped Adjacency (initialize with an adjacency matrix with self-loops).
[0083] In the second step, we use 60% high-confidence clustering samples with pseudo labels to optimize ,Right now Figure 4 Step 2:Refine with High-confidence Samples in the Refine with High-confidence Samples. Specifically, construct pseudo labels for these high-confidence samples based on cluster-ID, because high-confidence samples are always reliable, and further add edges (Added Edges) or remove edges (Removed Edges) to the adjacency matrix when the sample pairs have the same / different pseudo labels. Here, samples that are closer to the corresponding cluster center have higher confidence.
[0084] In this way, the proposed clustering refines the adjacency matrix It can better reveal the homogeneity and heterogeneity between different nodes. The homogeneity term encourages The homogeneity element of is equal to 1, which means that the embeddings of homogeneous nodes in two different views are pulled together, and the heterogeneity term makes The heterogeneous elements in are close to 0, thus pushing away the embeddings of heterogeneous nodes in the two views.
[0085] In one embodiment, Figure 3 As shown, a schematic flowchart of a step of reducing the correlation of a first augmented image and a second augmented image at a feature level is provided, and the specific steps are as follows:
[0086] Step 302: Obtain cluster-level feature maps according to the feature maps of the first augmented map and the second augmented map.
[0087] Step 304: Obtain a cross-view feature correlation matrix according to the cluster-level feature maps of the first augmented map and the second augmented map.
[0088] Step 306: Based on the cross-view feature correlation matrix and the identity matrix , construct the feature layer mean square error loss function.
[0089] Step 308, when the feature layer mean square error loss function is less than a preset threshold, a feature layer correlation reduction feature map is obtained.
[0090] In this implementation, when the feature layer mean square error loss function is less than a preset threshold, the cross-view feature correlation matrix is equal to the identity matrix ,The feature level correlation reduction mechanism proposed in the present ,invention reduces the redundant information in the learning representation, ,and thereby improves the discriminative ability of the features.
[0091] Specifically, by reading out the function Get cluster-level embeddings , read out function By dividing the samples into K groups and outputting the average value of each group, the cross-view feature correlation matrix is calculated. , the calculation formula of the cross-view feature correlation matrix is as follows:
[0092]
[0093] in, represents the cosine similarity between the i-th dimension feature of the first augmented view and the j-th dimension feature of the second augmented view.
[0094] In one embodiment, according to the cross-view feature correlation matrix and the identity matrix , constructing the feature layer mean square error loss function includes: according to the cross-view feature correlation matrix and the identity matrix , construct the feature layer mean square error loss function as:
[0095]
[0096] in, is the dimension of the learned latent embedding, , represents the cross-view feature correlation matrix, represents the cosine similarity between the i-th dimension feature of the first augmented view and the j-th dimension feature of the second augmented view.
[0097] In this embodiment, the first term means that the same dimensions of the learned features in the two enhanced views are forced to be consistent with each other, and the second term makes the different dimensions of the potential representation less correlated. In this way, the redundant information in the learned representation is reduced, thereby improving the discriminative ability of the features and avoiding the representation collapse problem.
[0098] In a specific embodiment, the objective function of constructing a dual correlation reduction network is , the objective function Including dual correlation reduces the loss of structure , reconstruction loss and clustering loss , the objective function It is expressed as:
[0099]
[0100] in, represents the weight, is the loss of DCRM, is the joint mean square error (MSE) reconstruction loss of node attributes and graph structure used in DFCN, including and , represents the KL divergence, i.e., the self-supervised clustering loss, using the objective function Training of dual correlation reduction networks.
[0101] In one embodiment, the step of obtaining a dual correlation reduction feature map includes: obtaining a dual correlation reduction feature map based on a sample-level correlation reduction feature map and a feature-level correlation reduction feature map.
[0102] In a specific embodiment, Figure 5As shown, a general flow chart of a graph deep clustering method based on dual correlation reduction is provided, and a data augmentation module (Graph Augmentation Module) is defined. In this module, a diffusion matrix is obtained by performing graph diffusion on the adjacency matrix (Adjacency Matrix) of the original target relationship graph, and a classification graph adjacency matrix is obtained by constructing a KNN graph adjacency matrix (KNN Graph) according to the relationship between the nodes in the original target relationship graph, and Gaussian noise is generated to form a feature perturbation matrix, and the feature perturbation is added to the classification graph adjacency matrix and the diffusion matrix respectively, and the first augmented graph and the second augmented graph obtained are respectively generated through an encoder (Siamese) to generate feature graphs in the form of node embeddings (Node Embeddings). By performing data augmentation on the original target relationship graph, the dual correlation reduction network can be guided to learn more discriminative embeddings from the two generated augmented graphs, and the feature graphs of the first augmented graph and the second augmented graph are respectively input into the DCRM module, and the generated sample-level and feature-level correlation reduction feature graphs are added (Addition) to form a dual correlation reduction feature graph, and the K-means algorithm and clustering guidance (Clustering Guidance) clusters the dual correlation reduction feature map and outputs the clustering map through the decoder. The data enhancement structure and the dual correlation reduction structure can solve the representation collapse problem in the field of deep graph clustering, thereby improving the clustering performance.
[0103] In a specific embodiment, Figure 6 As shown in the figure, a schematic diagram of GPU memory cost ablation of the graph deep clustering method based on dual correlation reduction on ACM (Association for Computing Machinery Digital Library), CITE (Citeseer) and DBLP (Digital Bibliography & Library Project) datasets is provided. Through the ablation study of GPU memory cost (GPU Memory Cost), it can be found that the method of the present invention (OURS) occupies much less memory space than MVGRL. The method proposed in the present invention does not generate negative samples, so it can save GPU memory costs.
[0104] It should be understood that although Figure 1-5 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 1-5 At 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.
[0105] In one embodiment, Figure 7 As shown, a graph deep clustering device based on dual correlation reduction is provided, including a graph acquisition module 702, a data enhancement module 704, a dual correlation reduction module 706 and a graph generation module 708, wherein:
[0106] A graph acquisition module 702, used to acquire an original target relationship graph to be clustered;
[0107] A data enhancement module 704 is used to perform data enhancement on the original target relationship graph based on feature similarity and graph diffusion, respectively, to obtain a classification graph adjacency matrix and a diffusion matrix, to obtain a first augmented graph based on the classification graph adjacency matrix and Gaussian noise, and to obtain a second augmented graph based on the diffusion matrix and Gaussian noise;
[0108] A dual correlation reduction module 706 is used to input the feature graphs of the first augmented graph and the second augmented graph into a pre-trained dual correlation reduction network to perform sample-level correlation reduction and feature-level correlation reduction, respectively, to obtain a dual correlation reduction feature graph;
[0109] The graph generation module 708 is used to perform graph clustering on the dual correlation reduction feature graph to obtain a target relationship clustering graph.
[0110] In one embodiment, the data enhancement module 704 is further used to calculate the cosine similarity between the nodes of the original target relationship graph to obtain an adjacency matrix based on similarity; and perform KNN algorithm classification on the adjacency matrix based on similarity to obtain a classified graph adjacency matrix:
[0111]
[0112] in, Represents the normalization function, represents the K-nearest neighbor algorithm, The feature matrix representing the original target relationship graph, Represents the transpose operation of a matrix.
[0113] In one embodiment, the data enhancement module 704 is further used to perform generalized graph diffusion on the adjacency matrix of the original target relationship graph to obtain a diffusion matrix:
[0114]
[0115] in, is the transmission probability, represents the identity matrix, , is the adjacency matrix of the original target relationship graph, is the normalized adjacency matrix, The feature matrix representing the original target relationship graph, Represents the matrix inversion operation.
[0116] In one of the embodiments, the dual correlation reduction module 706 is also used to obtain a cross-view sample correlation matrix based on the feature graphs of the first augmented graph and the second augmented graph; construct a sample-level mean square error loss function based on the cross-view sample correlation matrix and the clustering optimization adjacency matrix; when the sample-level mean square error loss function is less than a preset threshold, obtain a sample-level correlation reduction feature graph.
[0117] In one embodiment, the dual correlation reduction module 706 is further used to obtain cluster-level feature maps according to the feature maps of the first augmented map and the second augmented map, respectively; obtain a cross-view feature correlation matrix according to the cluster-level feature maps of the first augmented map and the second augmented map; and obtain a cross-view feature correlation matrix according to the cross-view feature correlation matrix and the unit matrix. , construct the mean square error loss function of the feature layer; when the mean square error loss function of the feature layer is less than the preset threshold, the correlation reduction feature map of the feature level is obtained.
[0118] In one embodiment, the dual correlation reduction module 706 is further used to construct a sample-level mean square error loss function according to the cross-view sample correlation matrix and the cluster optimization adjacency matrix:
[0119]
[0120] in, represents the cross-view sample correlation matrix, represents the clustering optimized adjacency matrix, , If is equal to 1, If is equal to 0, yes and The cosine similarity between and Denote the i-th node embedding of the first augmented view and the j-th node embedding of the second augmented view.
[0121] In one embodiment, the dual correlation reduction module 706 is further configured to reduce the cross-view feature correlation matrix and the identity matrix , construct the feature layer mean square error loss function as:
[0122]
[0123] in, is the dimension of the learned latent embedding, , represents the cross-view feature correlation matrix, represents the cosine similarity between the i-th dimension feature of the first augmented view and the j-th dimension feature of the second augmented view.
[0124] In one embodiment, 706 is further configured to obtain a dual correlation reduction feature map according to the sample-level correlation reduction feature map and the feature-level correlation reduction feature map.
[0125] For the specific limitations of the graph depth clustering device based on dual correlation reduction, please refer to the limitations of the graph depth clustering method based on dual correlation reduction in the above text, which will not be repeated here. Each module in the above-mentioned graph depth clustering device based on dual correlation reduction 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.
[0126] 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 8 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 depth clustering method based on dual correlation reduction 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 covering the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.
[0127] Those skilled in the art will understand that Figure 8The 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.
[0128] 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.
[0129] 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.
[0130] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented 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 may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of 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), etc.
[0131] The technical features of the above embodiments may be combined arbitrarily. 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.
[0132] 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 deep clustering method based on double correlation reduction, characterized in that: The method comprises: Obtain the original target relationship graph to be clustered; Based on feature similarity and graph diffusion, data enhancement is performed on the original target relationship graph to obtain a classification graph adjacency matrix and a diffusion matrix, a first augmented graph is obtained according to the classification graph adjacency matrix and Gaussian noise, and a second augmented graph is obtained according to the diffusion matrix and Gaussian noise; Inputting the feature graphs of the first augmented graph and the second augmented graph into a pre-trained dual correlation reduction network to perform sample-level correlation reduction and feature-level correlation reduction, respectively, to obtain a dual correlation reduction feature graph; Performing graph clustering on the dual correlation reduction feature graph to obtain a target relationship clustering graph; Based on the feature similarity, the original target relationship graph is data enhanced to obtain a classification graph adjacency matrix including: Calculating the cosine similarity between the nodes of the original target relationship graph to obtain an adjacency matrix based on similarity; The similarity-based adjacency matrix is classified using the KNN algorithm to obtain a classification graph adjacency matrix; the classification graph adjacency matrix is expressed as: ; in, Represents the normalization function, represents the K-nearest neighbor algorithm, The feature matrix representing the original target relationship graph, Represents the transpose operation of a matrix; Based on graph diffusion, data enhancement is performed on the original target relationship graph, and the obtained diffusion matrix includes: The adjacency matrix of the original target relationship graph is subjected to graph diffusion to obtain a diffusion matrix; the diffusion matrix is expressed as: ; in, is the transmission probability, represents the identity matrix, , is the adjacency matrix of the original target relationship graph, is the normalized adjacency matrix, The feature matrix representing the original target relationship graph, Represents the matrix inversion operation.
2. The method according to claim 1, characterized in that The step of inputting the feature graphs of the first augmented graph and the second augmented graph into a pre-trained dual correlation reduction network to perform sample-level correlation reduction comprises: Obtaining a cross-view sample correlation matrix according to the feature graphs of the first augmented graph and the second augmented graph; Constructing a sample-level mean square error loss function according to the cross-view sample correlation matrix and the cluster-optimized adjacency matrix; When the sample layer mean square error loss function is less than a preset threshold, a sample layer correlation reduction feature map is obtained.
3. The method according to claim 1, characterized in that The step of inputting the feature graphs of the first augmented graph and the second augmented graph into a pre-trained dual correlation reduction network to reduce the correlation at the feature level comprises: Obtaining cluster-level feature maps respectively according to the feature maps of the first augmented map and the second augmented map; Obtaining a cross-view feature correlation matrix according to the cluster-level feature maps of the first augmented image and the second augmented image; According to the cross-view feature correlation matrix and the identity matrix , construct the feature layer mean square error loss function; When the feature layer mean square error loss function is less than a preset threshold, a feature layer correlation reduction feature map is obtained.
4. The method according to claim 2, characterized in that: The constructing of a sample layer mean square error loss function according to the cross-view sample correlation matrix and the cluster optimization adjacency matrix comprises: According to the cross-view sample correlation matrix and cluster optimization adjacency matrix, the sample layer mean square error loss function is constructed as follows: ; in, represents the cross-view sample correlation matrix, represents the clustering optimized adjacency matrix, , If is equal to 1, If is equal to 0, yes and The cosine similarity between and Denote the i-th node embedding of the first augmented view and the j-th node embedding of the second augmented view.
5. The method according to claim 3, characterized in that: The cross-view feature correlation matrix and the identity matrix , constructing the feature layer mean square error loss function includes: According to the cross-view feature correlation matrix and the identity matrix , construct the feature layer mean square error loss function as: ; in, is the dimension of the learned latent embedding, , represents the cross-view feature correlation matrix, represents the cosine similarity between the i-th dimension feature of the first augmented view and the j-th dimension feature of the second augmented view.
6. The method according to any one of claims 2 to 5, characterized in that: The steps of obtaining the dual correlation reduction feature map include: A dual correlation reduction feature map is obtained based on the sample-level correlation reduction feature map and the feature-level correlation reduction feature map.
7. A graph deep clustering device based on dual correlation reduction, characterized in that: The device comprises: A graph acquisition module, used to obtain the original target relationship graph to be clustered; A data enhancement module, used to perform data enhancement on the original target relationship graph based on feature similarity and graph diffusion, respectively, to obtain a classification graph adjacency matrix and a diffusion matrix, to obtain a first augmented graph according to the classification graph adjacency matrix and Gaussian noise, and to obtain a second augmented graph according to the diffusion matrix and Gaussian noise; A dual correlation reduction module, used for inputting the feature graphs of the first augmented graph and the second augmented graph into a pre-trained dual correlation reduction network to perform sample-level correlation reduction and feature-level correlation reduction, respectively, to obtain a dual correlation reduction feature graph; A graph generation module, used for performing graph clustering on the dual correlation reduction feature graph to obtain a target relationship clustering graph; The data enhancement module is also used to calculate the cosine similarity between the nodes of the original target relationship graph to obtain an adjacency matrix based on similarity; the adjacency matrix based on similarity is classified by the KNN algorithm to obtain a classification graph adjacency matrix; the classification graph adjacency matrix is expressed as: ; in, Represents the normalization function, represents the K-nearest neighbor algorithm, The feature matrix representing the original target relationship graph, Represents the transpose operation of a matrix; The data enhancement module is also used to perform graph diffusion on the adjacency matrix of the original target relationship graph to obtain a diffusion matrix; the diffusion matrix is expressed as: ; in, is the transmission probability, represents the identity matrix, , is the adjacency matrix of the original target relationship graph, is the normalized adjacency matrix, The feature matrix representing the original target relationship graph, Represents the matrix inversion operation.
8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.