Structural multi-view clustering method based on heterogeneous graph random walk

Abstract

The invention discloses a structural multi-view clustering method based on heterogeneous graph random walk, and belongs to the field of natural language processing and machine learning. The invention provides a structural multi-view clustering method based on heterogeneous graph random walk. The structural multi-view clustering method comprises the following steps: firstly, constructing a multi-view heterogeneous graph comprising sample nodes and view nodes; a multi-step random walk strategy is designed on the constructed multi-view heterogeneous graph, a cross-view high-order sample structure is found, and meanwhile, a structure specific to the view is reserved; wherein the random walk mainly comprises a multi-view single-step random walk process and a multi-view multi-step random walk process; and finally, constructing a structured multi-view clustering network, and performing intra-view and cross-view mixed structure learning by taking a unified sample-level structure generated by multi-step random walk as guidance. According to the method, the representation capability of the multi-view data is enhanced, and the clustering performance is improved.

Description

Technical Field

[0001] This invention belongs to the fields of natural language processing and machine learning technology, and specifically relates to a structural multi-view clustering method based on heterogeneous graph random walks. Background Technology

[0002] Multi-view data refers to a dataset that describes the same data sample from different perspectives. It includes multiple data sources or feature sets, collectively achieving a comprehensive description of the same data sample. In the biological field, multi-view data can include different omics (such as DNA methylation, RNA expression) information of the same cancer subtype; in the text field, it can include information describing the news topic in different views of the same news (such as headlines, content).

[0003] Multi-view clustering utilizes multi-view data to divide data samples into disjoint clusters. Existing multi-view clustering methods mostly focus on data representation learning, often neglecting data structure information. The structure between data samples can reveal hidden similarities and associations beyond the representation space, which is valuable for representation learning and can promote the discovery of potential patterns in the data. Some methods have attempted to explore the intrinsic relationships between data samples using structural information: GDMVC uses a graph convolutional autoencoder to combine features and structural information to capture underlying data relationships; DIVIDE introduces high-order random walks in each view to learn latent structural information to guide multi-view representation learning and clustering. These methods learn representations by combining view-specific structural information, but they do not achieve unified sample-level data partitioning.

[0004] In multi-view scenarios, data samples are described by different views, characterized by multi-view representations rather than explicit sample representations. In multi-view clustering tasks, only the view structure can be used to assist in obtaining sample data partitions; the sample structure cannot be used directly. Significant differences exist between multi-view structures, making it difficult to form a consistent sample structure across views, resulting in representation failure. Each view has its own specific structure, lacking mutual support, making it impossible to determine the contribution weight of each view structure to the sample structure, and making it difficult to guarantee the optimality of the sample structure. Summary of the Invention

[0005] To address the technical problems of existing multi-view clustering methods, this invention provides a structural multi-view clustering method based on heterogeneous graph random walks.

[0006] The technical solution adopted by this invention to solve the technical problem is as follows:

[0007] This invention provides a structural multi-view clustering method based on heterogeneous graph random walks, comprising the following steps:

[0008] Step 1: Construct a multi-view heterogeneous graph containing sample nodes and view nodes;

[0009] Step 2: Perform a multi-step random walk on the multi-view heterogeneous graph; the multi-step random walk includes a multi-view single-step random walk and a multi-view multi-step random walk;

[0010] Step 3: Construct a structured multi-view clustering network and use the unified sample-level structure generated by multi-step random walks as a guide to learn mixed structures within and across views.

[0011] Furthermore, in the multi-view heterogeneous graph, each sample node is linked to multiple view nodes that describe the sample with the same probability, and there are no links between sample nodes; each view node is linked to other view nodes according to its corresponding transition probability; each view node should be linked to its sample node.

[0012] Furthermore, in the multi-view heterogeneous graph, the structural relationships between sample nodes are explored through the edge weights between view nodes; the edge weights are defined using Gaussian kernel similarity.

[0013] Furthermore, in the multi-view single-step random walk, an initial transformation from sample nodes to view nodes is performed first, then a transformation between view nodes is performed, and finally a transformation from view nodes back to sample nodes is performed.

[0014] Furthermore, in the initial transformation process from the sample node to the view node, from the sample node... To begin, randomly select a view index. Corresponding to a view node The probability of selecting each view node follows a uniform distribution; during the transition between view nodes, for each sample node... Given the first A view, from the view node Switch to another view node in the same view The transition probability is determined by the first... Transition matrix of each view Defined; during the transformation process from the view node back to the sample node, the data reaches the view node. Then, return to the corresponding sample node. The probability is 1.

[0015] Furthermore, in the multi-view multi-step random walk, a view is randomly selected in each step, and a multi-view single-step random walk is performed.

[0016] Furthermore, in the In the step, from the sample node to sample node The transition probability matrix is ​​denoted as :

[0017]

[0018] in, Indicates the first The set of view nodes selected from all sample nodes in the step. Indicates the number of view nodes. Indicates the first Step to the first The expected transition probability matrix of each view node. Represents the sequence of all possible view node selections ( , , ..., The transition probability matrix follows a uniform distribution. Represented as:

[0019]

[0020] in, Indicates the number of transition steps. Indicates in The total number of all selectable view nodes in the step.

[0021] Furthermore, the multi-view clustering network includes a twin view encoder, a momentum encoder, and a cross-view predictor; the twin view encoder and momentum encoder are used to learn the semantic representation of the view; the cross-view predictor projects the representation of one view into the representation space of other views by applying a two-layer nonlinear fully connected network.

[0022] Furthermore, in the multi-view clustering network, the overall objective function is:

[0023]

[0024] In-view structure learning loss The calculation formula is:

[0025]

[0026] Cross-view structure learning loss The calculation formula is:

[0027]

[0028] in, Indicates the number of view nodes; Indicates the view index; Represents cross-entropy; This represents a pseudo-target that contains sample relationships between multiple views. Indicates size is Matrix; express Similarity; This indicates the hidden layer embedding of the twin view encoder; This indicates the hidden layer embedding of the momentum encoder; Indicates the first The view is mapped to the first Hidden layer embedding of each view Indicates the first A view Embedded.

[0029] Furthermore, the in-view structure learning loss and cross-view structure learning loss The equations establish pseudo-targets as a unified sample-level structure guide:

[0030]

[0031] in, =0.5, Indicates the process Obtained by random walk The normalized matrix, Representing normalized multi-view Nearest neighbor matrix.

[0032] The beneficial effects of this invention are:

[0033] This invention provides a structured multi-view clustering method based on heterogeneous graph random walks. First, a multi-view heterogeneous graph containing sample nodes and view nodes is constructed. A multi-step random walk strategy is designed on the constructed multi-view heterogeneous graph to discover higher-order sample structures across views while preserving view-specific structures. The random walks mainly include multi-view single-step random walks and multi-view multi-step random walks. Finally, a structured multi-view clustering network is constructed, and the unified sample-level structure generated by the multi-step random walks is used as a guide for learning mixed intra-view and cross-view structures, enhancing the representation capability of multi-view data and improving clustering performance. Attached Figure Description

[0034] Figure 1 This is a schematic diagram illustrating the implementation principle of a structured multi-view clustering network developed by multi-step random walks of multi-view heterogeneous graphs according to the present invention. Detailed Implementation

[0035] The present invention will be further described in detail below with reference to the accompanying drawings.

[0036] This invention provides a structural multi-view clustering method based on heterogeneous graph random walks, which mainly includes steps such as constructing multi-view heterogeneous graphs, heterogeneous random walks, and hybrid structure learning. The specific implementation process is as follows:

[0037] Step 1: Constructing a multi-view heterogeneous graph;

[0038] To obtain the structural information required for multi-view clustering, this invention constructs a multi-view heterogeneous graph based on multi-view data, which contains two types of nodes: sample nodes and view nodes. Specifically, sample nodes are not explicitly represented and can only be linked to view nodes; view nodes are represented using the embedding of hidden layers and can be linked to both sample nodes and view nodes simultaneously.

[0039] For each sample node, it is linked to a node that describes the sample with the same probability. There are 3 view nodes, and there are no links between sample nodes. For each view node, it is linked to other view nodes according to its corresponding transition probability; and each view node must be linked to its sample nodes. Based on this, the structural relationships between sample nodes can be explored through the edge weights between view nodes.

[0040] Consider a containing A dataset of samples, consisting of a set of sample nodes. Represented. Each sample node in it. and A set of view nodes is associated with several different view nodes. It means that, among them Indicates the relationship with sample nodes The associated first One view node; .

[0041] This invention constructs a transition matrix for each view. More specifically, it first constructs a fully connected adjacency matrix using intra-batch samples of each view, treats embeddings as nodes, and defines edge weights between view nodes using the following Gaussian kernel similarity:

[0042]

[0043] in, Indicates the first Within the first view The first view node and the first Similarity between view nodes Indicates the first The first view in the An embedding, Indicates the first The first view in the An embedding, where the superscript T indicates transpose. Indicates bandwidth parameter, =0.1. Next, by analyzing the fully connected adjacency matrix... Perform row-wise normalization to obtain the first row. Transition matrix of each view .

[0044] A multi-step random walk is designed on the constructed multi-view heterogeneous graph, and its specific implementation process is as follows:

[0045] S2.1: Multi-view single-step random walk process;

[0046] S2.1.1: Initial transformation from sample node to view node: from sample node To begin, randomly select a view index. Corresponding to a view node The probability of selecting each view node follows a uniform distribution:

[0047]

[0048] in, Represents sample nodes Select view node The probability of.

[0049] S2.1.2: Transitions between view nodes: For each sample node Given the selected first A view, from the view node Switch to another view node in the same view The transition probability is determined by the first... Transition matrix of each view Define:

[0050]

[0051] in, Indicates that in the known number of... Transition matrix of each view Under the premise that view nodes Select view node The probability, Indicates the first View nodes in a view With view node The similarity.

[0052] S2.1.3: Transition from view node back to sample node: Reaching the view node Then, return to the corresponding sample node. The probability is 1:

[0053]

[0054] Therefore, for a multi-view single-step random walk, starting from the sample node to sample node The transition probability is:

[0055]

[0056] S2.2: Multi-view, multi-step random walk process;

[0057] A multi-view, multi-step random walk can be viewed as an iteration of the aforementioned multi-view, single-step random walk process. At each step... In the middle, randomly select a view node. And perform the above-described multi-view single-step random walk. For the multi-view multi-step random walk process, considering the independence and randomness of the view selection at each step, from the sample node... to sample node The The step transition probability matrix is ​​denoted as , can be represented as:

[0058]

[0059] in, Indicates the first The view node selection sequence set for all sample nodes in the step. Indicates the first Step to the first The expected transition probability matrix of each view node. Represents the sequence of all possible view node selections ( , , ..., They follow a uniform distribution.

[0060] Therefore, the first Step transition probability matrix It can be represented as:

[0061]

[0062] Where t represents the number of transition steps, Indicates in The number of all selectable view nodes in the step.

[0063] This invention achieves a complete heterogeneous graph of walking paths through multi-step random walks. It integrates structural information from multiple perspectives, captures complex relationships, and enhances the data's representational capabilities. The resulting transition probability matrix provides unified sample-level structural guidance for multi-view clustering at the sample level.

[0064] This invention develops a structured multi-view clustering network through multi-step random walks of multi-view heterogeneous graphs. This network is a lightweight network guided by a unified sample-level structure and is achieved through hybrid structure learning within and across views.

[0065] The multi-view clustering network proposed in this invention mainly includes multiple twin-view encoders, multiple momentum encoders, and multiple cross-view predictors. The twin-view encoders and momentum encoders are used to learn the semantic representations of the views. The cross-view predictors project the representation of one view into the representation space of other views by applying a two-layer nonlinear fully connected network.

[0066] Specifically, this will be illustrated using two twin-view encoders, two momentum encoders, and two cross-view predictors as examples, such as... Figure 1 As shown, the vision is learned through the twin view encoder 1 and the momentum encoder 1 respectively. Figure 1 Semantic representation of the content, generating hidden layer embedding 1 and hidden layer embedding 2 respectively. The hidden layer embedding 1 input to the cross-view predictor 1 will view Figure 1 The representation of is projected onto the representation space of view 2 to obtain the view. Figure 1 Mapping to View 2's embedding; embedding the hidden layer Pseudo-targets are obtained by constructing multi-view heterogeneous graphs, heterogeneous random walks, and hybrid structure learning, and then embedding the hidden layers into layers 1 and 2. Together with pseudo-targets, we learn the in-view structure;

[0067] Semantic representations of the content of view 2 are learned using twin view encoder 2 and momentum encoder 2, respectively, and hidden layer embedding 2 and hidden layer embedding 2 are generated. The hidden layer embedding 2 inputs to the cross-view predictor 2, projecting the representation of view 2 onto the view. Figure 1 In the representation space, view 2 is mapped to view... Figure 1 Embedding; embedding the hidden layer By constructing a multi-view heterogeneous graph, heterogeneous random walk, and hybrid structure learning to obtain pseudo-targets, the hidden layer is embedded in layer 2. Together with pseudo-targets, we learn the in-view structure;

[0068] Embed the hidden layer False targets and visual Figure 1The embeddings mapped to View 2 are used together to learn the cross-view structure and embed the hidden layer. False targets and View 2 mapping to view Figure 1 The embedding is used to learn the cross-view structure.

[0069] Specifically, the overall objective function of this invention is as follows:

[0070]

[0071] in, Represents the overall objective function. and These represent the learning loss for the structure within the view and the learning loss for the structure across the view, respectively.

[0072] set up Indicates the first The encoder parameters for each view. This indicates the hidden layer embedding of the twin view encoder. This indicates the hidden layer embedding of the momentum encoder. Indicates the first The view is mapped to the first Hidden layer embedding of a view, Indicates the first The embedding of each view results in a loss for learning the structure within the view. The calculation formula is as follows:

[0073]

[0074] in, Represents cross-entropy. This represents a pseudo-target that contains sample relationships between multiple views. Indicates size is , express The similarity, after being processed by row normalization, yields the following parameters:

[0075]

[0076] in, This indicates similarity calculation. Indicates the first Hidden layer embeddings for each sample, express Hidden layer embeddings for each sample, =0.5.

[0077] set up Indicates the first The parameters of the cross-view predictor for each view are then used to calculate the cross-view structure learning loss. The calculation formula is as follows:

[0078]

[0079] in, Both represent view indexes. Indicates the first The view is mapped to the first Embedding of a view.

[0080] To achieve unified structure learning across multiple views, by integrating multiple views... Figure 1 Order and higher order ( (Rank) structural information, in-view structural learning loss and cross-view structure learning loss A pseudo-target was established in the equation. As a unified sample-level structure guide, namely:

[0081]

[0082] in, =0.5, Indicates the process Obtained by random walk The normalized matrix, This indicates a normalized multi-view. The nearest neighbor matrix is ​​defined as follows:

[0083]

[0084] in, Indicates matrix normalization; This represents the adjacency matrix after matrix filtering. This represents a matrix filter used to determine the characteristics of each sample. Find the nearest neighbors and preserve their weights. Specifically, if the... The first sample and the first The sample at the th If each element is considered a neighbor pair in a view, then each element... Set as Otherwise, each element It was set to 0.

[0085] Compared with existing technologies, the structural multi-view clustering method based on heterogeneous graph random walk proposed in this invention establishes a unified sample-level structure. Extensive experiments on five real datasets demonstrate that the structural multi-view clustering method based on heterogeneous graph random walk of this invention outperforms existing multi-view clustering models (GCFAGG, CVCL, MAGA, and SURER) in clustering performance.

[0086] The five real datasets used in this invention are Scene15, UCI-Digit, CUB, Prokaryotic, and ALOI. Details of the datasets are shown in Table 1.

[0087] Table 1

[0088]

[0089] In this invention, clustering accuracy (ACC), normalized mutual information (NMI), and adjusted RAND index (ARI) are used as evaluation metrics for clustering results. Higher values ​​for ACC, NMI, and ARI indicate better clustering performance.

[0090] This invention is based on several deep multi-view clustering algorithms. Specifically, these deep multi-view clustering algorithms can be broadly classified into two categories: representation-based multi-view clustering models and structure-based multi-view clustering models. The representation-based multi-view clustering models include GCFAGG and CVCL; the structure-based multi-view clustering models include MAGA and SURER.

[0091] The experimental results are shown in Table 2. "-" means that the present invention cannot obtain clustering results on this dataset because large-scale datasets consume a lot of computing resources.

[0092] Table 2. Model comparison experiments on five real datasets (%)

[0093]

[0094] The experimental results above show that, on five real datasets (Scene15, UCI-Digit, CUB, Prokaryotic, and ALOI), the evaluation metrics ACC, NMI, and ARI of the proposed method are all higher than those of existing multi-view clustering models (GCFAGG, CVCL, MAGA, and SURER). This indicates that the proposed method enhances the representation capability of multi-view data and improves clustering performance.

[0095] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A structural multi-view clustering method based on heterogeneous graph random walks, characterized in that, Includes the following steps: Step 1: Construct a multi-view heterogeneous graph containing sample nodes and view nodes; Step 2: Perform a multi-step random walk on the multi-view heterogeneous graph; the multi-step random walk includes a multi-view single-step random walk and a multi-view multi-step random walk; Step 3: Construct a structured multi-view clustering network and use the unified sample-level structure generated by multi-step random walks as a guide to learn mixed structures within and across views; The multi-view clustering network includes a twin view encoder, a momentum encoder, and a cross-view predictor; the twin view encoder and momentum encoder are used to learn the semantic representation of the view; the cross-view predictor projects the representation of one view into the representation space of other views by applying a two-layer nonlinear fully connected network.

2. The structural multi-view clustering method based on heterogeneous graph random walks according to claim 1, characterized in that, In the multi-view heterogeneous graph, each sample node is linked to multiple view nodes that describe the sample with the same probability, and there are no links between sample nodes. Each view node is linked to other view nodes according to its corresponding transition probability; each view node should be linked to its sample node.

3. The structural multi-view clustering method based on heterogeneous graph random walks according to claim 1, characterized in that, In the multi-view heterogeneous graph, the structural relationship between sample nodes is explored by the edge weights between view nodes; The edge weights are defined using Gaussian kernel similarity.

4. The structural multi-view clustering method based on heterogeneous graph random walks according to claim 1, characterized in that, In the multi-view single-step random walk, an initial transformation from sample nodes to view nodes is performed first, then a transformation between view nodes is performed, and finally a transformation from view nodes back to sample nodes is performed.

5. The structural multi-view clustering method based on heterogeneous graph random walks according to claim 4, characterized in that, During the initial transformation from the sample node to the view node, from the sample node To begin, randomly select a view index. Corresponding to a view node The probability of selecting each view node follows a uniform distribution; during the transition between view nodes, for each sample node... Given the first A view, from the view node Switch to another view node in the same view The transition probability is determined by the first... Transition matrix of each view Defined; during the transformation process from the view node back to the sample node, the data reaches the view node. Then, return to the corresponding sample node. The probability is 1.

6. The structural multi-view clustering method based on heterogeneous graph random walks according to claim 1, characterized in that, In the multi-view multi-step random walk, a view is randomly selected in each step, and a multi-view single-step random walk is performed.

7. The structural multi-view clustering method based on heterogeneous graph random walks according to claim 6, characterized in that, In the In the step, from the sample node to sample node The transition probability matrix is ​​denoted as : ， in, Indicates the first The set of view nodes selected from all sample nodes in the step. Indicates the number of view nodes. Indicates the first Step to the first The expected transition probability matrix of each view node. Represents the sequence of all possible view node selections ( , , ..., The transition probability matrix follows a uniform distribution. Represented as: ， in, Indicates the number of transition steps. Indicates in The total number of all selectable view nodes in the step.

8. The structural multi-view clustering method based on heterogeneous graph random walks according to claim 1, characterized in that, In the multi-view clustering network, the overall objective function is: ， In-view structure learning loss The calculation formula is: ， Cross-view structure learning loss The calculation formula is: ， in, Indicates the number of view nodes; Indicates the view index; Represents cross-entropy; This represents a pseudo-target that contains sample relationships between multiple views. Indicates size is Matrix; express Similarity; This indicates the hidden layer embedding of the twin view encoder; This indicates the hidden layer embedding of the momentum encoder; Indicates the first The view is mapped to the first Hidden layer embedding of a view; Indicates the first Embedding of a view.

9. A structural multi-view clustering method based on heterogeneous graph random walks according to claim 8, characterized in that, The in-view structure learning loss and cross-view structure learning loss The equations establish pseudo-targets as a unified sample-level structure guide: ， in, =0.5, Indicates the process Obtained by random walk The normalized matrix, Representing normalized multi-view Nearest neighbor matrix.

Images