Multi-view contrast clustering method based on granular balls

Abstract

The invention discloses a multi-view contrast clustering method based on a granular ball, and the method comprises the following steps: 1, obtaining multi-view data, and forming a data set; 2, constructing a multi-view comparison clustering model, wherein the multi-view comparison clustering model comprises a data processing module, a particle ball generation module and an output module; step 3, training the multi-view contrast clustering model to obtain a trained multi-view contrast clustering model; 4, comparing the clustering model according to the trained multiple views to obtain a required clustering result; the method comprises the following steps: segmenting a sample set into coarse-grained pellets, and establishing association between in-view pellets and cross-view pellets; the associations are enhanced in a shared potential space, so that multi-granularity comparative learning is realized, the particle balls are located between an instance level and a cluster level, and a local topological structure of a sample set is reserved.

Description

Technical Field

[0001] The present invention relates to the technical field of clustering methods, and in particular to a multi-view comparison clustering method based on a particle sphere. Background Art

[0002] Multi-view data is collected from different sensors or feature extractors and usually exhibits heterogeneity. For example, a web page contains pictures, texts, videos, etc., which are regarded as different views and reflect part of the information of the web page from their own perspectives. Multi-view clustering has received continuous attention in recent years, aiming to cluster multi-view data into multiple clusters in an unsupervised paradigm. The key challenge is how to balance the consistency and complementarity of multiple views so as to learn the most comprehensive consensus representation. Traditional multi-view clustering methods can be mainly divided into three categories: subspace learning, graph learning, and multi-kernel learning. These methods involve matrix decomposition and fusion, have high computational complexity, and are therefore difficult to apply to large-scale datasets, which limits their practical application.

[0003] In recent years, multi-view clustering methods based on deep learning have attracted widespread attention due to their excellent feature representation capabilities. Such methods extend deep single-view clustering methods, and select specific feature extractors for different views based on view characteristics. For example, the method proposed by Deep canonical correlation analysis uses a deep neural network to project the data of two views into a common space, where the feature representations of the two views are highly correlated, making it a nonlinear extension of canonical correlation analysis. The method proposed by Deep subspace clustering with sparsity prior is a deep subspace clustering method with sparse prior, which projects the input data into the latent space, maintains the local structure by minimizing the reconstruction loss, and introduces the sparse prior information into the latent representation learning to retain the sparse reconstruction relationship of the entire dataset. Duet Robust Deep Subspace Clustering uses two regularization specifications to model the collapse mode (such as intra-view noise) from the perspective of data reconstruction and latent self-expression, and also proposes a smoothing method for non-differentiable norms to optimize the loss function using a gradient-based method. DIMC-net: Deep Incomplete Multi-viewClustering Network extracts high-level features from multiple views through view-specific autoencoders and introduces constraints based on fusion graphs to preserve the local geometric structure of the data. Deep multi-view spectral clustering viaensemble applies ensemble clustering to fuse similarity graphs from different views and uses graph autoencoders to learn common spectral embeddings. It designs a unified optimization framework to simultaneously minimize graph reconstruction loss, orthogonality loss, and graph contrast loss. These methods combine deep learning with traditional multi-view learning ideas by introducing constraints from traditional methods (such as neighborhood graph constraints and self-expression constraints) into the latent space of deep module projections. This allows the model to learn concise and comprehensive representations that maximize the retention of structural information of the input data.

[0004] Existing multi-view contrastive learning methods usually operate at two scales: instance level and cluster level. Instance-level methods construct positive and negative pairs based on sample correspondences, aiming to bring positive pairs closer and push negative pairs further apart in the latent space. Cluster-level methods focus on computing cluster assignments for samples under each view and maximizing view consensus by reducing distribution differences, such as minimizing KL divergence or maximizing mutual information. However, both types of methods either introduce false negative pairs, resulting in reduced model discriminability, or ignore local structures and cannot explicitly measure the relationship between cross-view clusters. Summary of the invention

[0005] The present invention aims at solving the problems in the prior art and provides a multi-view comparison clustering method based on a sphere.

[0006] The technical solution adopted by the present invention is: a multi-view comparison clustering method based on particle spheres, comprising the following steps:

[0007] Step 1: Obtain multi-view data to form a data set;

[0008] Step 2: construct a multi-view comparison clustering model, which includes a data processing module, a particle sphere generation module and an output module;

[0009] The data processing module is used to reduce the dimension of the multi-view data in the data set to obtain the multi-view data features;

[0010] The particle sphere generation module is used to construct a particle sphere set according to each view corresponding to the multi-view data features, and to construct a particle sphere overlap relationship matrix within each view and a particle sphere correlation matrix between views;

[0011] The mask matrix is ​​obtained according to the particle-ball overlap relationship matrix within the view and the particle-ball correlation matrix between views, and the fusion feature is obtained through contrast learning;

[0012] The output module is used to output clustering results based on fusion features;

[0013] Step 3: Train the multi-view comparative clustering model to obtain a trained multi-view comparative clustering model;

[0014] Step 4: Obtain the desired clustering results based on the trained multi-view comparative clustering model.

[0015] Furthermore, the process of constructing the particle sphere set in step 2 is as follows:

[0016] In each view, k clusters, i.e., k spheres, are obtained by the k-means clustering method. The k spheres constitute a sphere set. The above process is repeated for each view to obtain the entire sphere set.

[0017] Furthermore, the center c of the sphere i and radius r i The calculation method is as follows:

[0018]

[0019] Where: i is the number of the ball, n i is the number of samples in the sphere, j is the serial number of the samples contained in the sphere, x j is the jth sample of the current ball.

[0020] Furthermore, the overlapping relationship matrix A v Elements in The following relations are satisfied:

[0021]

[0022] Where: v is the view number;

[0023] The distance between the particles is calculated as follows:

[0024]

[0025] Where: is the distance between the i-th sphere and the j-th sphere in the v-th view, is the center of the i-th sphere in the v-th view, is the center of the jth sphere in the vth view.

[0026] Furthermore, the inter-view particle-sphere correlation matrix P (m,n) Elements in satisfy:

[0027]

[0028] Where: m and n are view numbers, t i is a ball in view m, t j is a ball in view n, t both is the number of samples in the intersection of two spheres, and τ is the threshold parameter;

[0029] in:

[0030] t both =length(Id both )

[0031] Where: Id both is the common sample set of the i-th sphere in view m and the j-th sphere in view n.

[0032] Furthermore, the mask matrix M is:

[0033]

[0034] Where: P (n,m) P (m,n) The transpose of .

[0035] Furthermore, the loss function in the training process includes contrast loss and encoding loss;

[0036] L=L con +λL rec

[0037] Where: L is the loss function, L con is the contrast loss, Lrec is the encoding loss and λ is a parameter.

[0038] Furthermore, the contrast loss function is as follows:

[0039]

[0040] Where: V is the number of views, L (m,n) is the loss between views m and n.

[0041] Furthermore, the coding loss is:

[0042]

[0043] Where: E v (·;θ v ) is the encoder, D v (·;φ v ) is a decoder, is, v is the parameter, φ v is a parameter, N is the number of samples in the view, V is the number of views, v is the view number, and i is the sample number.

[0044] A system based on a multi-view comparison clustering method of a particle sphere, comprising a data acquisition and processing module, a particle sphere generation module and an output module;

[0045] The data acquisition and processing module is used to acquire a multi-view data set, perform dimensionality reduction on the multi-view data in the data set, and obtain multi-view data features;

[0046] The particle sphere generation module is used to construct a particle sphere set according to each view corresponding to the multi-view data features, and to construct a particle sphere overlap relationship matrix within each view and a particle sphere correlation matrix between views;

[0047] The mask matrix is ​​obtained according to the particle-ball overlap relationship matrix within the view and the particle-ball correlation matrix between views, and the fusion feature is obtained through contrast learning;

[0048] The output module is used to output clustering results based on fusion features.

[0049] The beneficial effects of the present invention are:

[0050] (1) The proposed method divides the sample set into coarse-grained spheres and establishes associations between the spheres within and across views. These associations are strengthened in a shared latent space, thereby achieving multi-granularity contrastive learning.

[0051] (2) The spheres in the present invention are located between instances and clusters, preserving the local topological structure of the sample set. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a structural schematic diagram of the multi-view comparison clustering model in the present invention.

[0053] Figure 2 This is the clustering result visualized by the method of the present invention using t-SNE on the MNIST-USPS dataset in an embodiment of the present invention.

[0054] Figure 3 In the embodiment of the present invention, Caltech 101-20 ( Figure 3 a) and Cora ( Figure 3 b) Parameter analysis experiment with different p and d, the vertical axis represents the clustering accuracy. DETAILED DESCRIPTION

[0055] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0056] A multi-view contrast clustering method based on particle spheres comprises the following steps:

[0057] Step 1: Obtain multi-view data to form a data set;

[0058] Given a multi-view dataset with N samples Each sample has V instances from different views, let d v is the feature dimension of the vth view, which usually has different values ​​in different views.

[0059] Step 2: Construct a multi-view comparison clustering model. The multi-view comparison clustering model includes a data processing module, a particle sphere generation module and an output module. The structure is as follows: Figure 1 shown.

[0060] The data processing module is used to reduce the dimension of the multi-view data in the data set to obtain the multi-view data features;

[0061] For subsequent comparison and fusion, the cross-view feature dimensions need to be unified, that is, different views are projected and transformed to the same dimension by the encoder. Deep autoencoders are used as representation learning frameworks to effectively extract basic low-dimensional embeddings from raw features.

[0062] For the vth view, E v (·;θ v ) is the encoder, D v (·;φ v ) is the decoder, θ v is the parameter, φ v As a parameter; The output dimension of is set to d. The feature matrix is ​​obtained Stored in the latent space, this feature is subsequently used to construct the particle sphere.

[0063] The features corresponding to the i-th instance of the v-th view It is expressed as:

[0064]

[0065] Where: For X v The i-th instance of .

[0066] The particle sphere generation module is used to construct a particle sphere set according to each view corresponding to the multi-view data features, and to construct a particle sphere overlap relationship matrix within each view and a particle sphere correlation matrix between views;

[0067] First, a granularity control parameter p is introduced, which can roughly reflect the granularity of the generated spheres. Let k be the total number of spheres generated under a single view. The number of samples N and k satisfy:

[0068]

[0069] Then, k-means is run in a single view to divide the image into k clusters, each of which is regarded as a sphere.

[0070] Then the i-th ball is represented by GB i ; The center of the sphere c i and radius r i The calculation method is as follows:

[0071]

[0072] Where: i is the number of the ball, n i is the number of samples in the sphere, j is the serial number of the samples contained in the sphere, x j is the jth sample of the current ball.

[0073] For different views, the same granularity parameter p is set, and the corresponding k is also the same, that is, each view is divided into the same number of granular balls.

[0074] For the vth view, H v Indicates that after encoder E v After projection, the above-mentioned sphere method is used to obtain the sphere set of the vth view make represents the union of the sphere sets of all views, C v R represents the center matrix of the vth view. v Represents the radius matrix of the vth view. For the i-th particle ball under the v-th view Its center is The radius is The calculation of the center and radius is gradient preserving. Let D vrepresents the particle-sphere distance matrix of the vth view. The distance between the i-th particle-sphere and the j-th particle-sphere is:

[0075]

[0076] Where: is the distance between the i-th sphere and the j-th sphere in the v-th view, is the center of the i-th sphere in the v-th view, is the center of the jth sphere in the vth view.

[0077] Based on D v The particle-ball overlap relationship matrix A in the view can be obtained v :Each of its elements satisfies the following relationship:

[0078]

[0079] Where: v is the view number;

[0080] Among them A v As part of the contrastive learning of positive and negative sample pairs, the matrix set Establish the relationship between particles and balls in the view.

[0081] The mask matrix is ​​obtained according to the particle-ball overlap relationship matrix within the view and the particle-ball correlation matrix between views, and the fusion feature is obtained through contrast learning;

[0082] Let's build the association between the cross-view instances:

[0083] Two balls in different views are considered neighbors in the latent space if each of them contains instances of the same samples from their respective views. However, this approach lacks robustness. When p is relatively large, the spheres also become large, and due to randomness, two cross-view balls may contain very few common samples.

[0084] The present invention solves this problem by the following method:

[0085] First, identify the common sample set between the two balls based on the stored sample indices:

[0086]

[0087] Where: and are two spheres of view m and n, each containing t i and t j samples, and Id(·) is the index of the sample contained in the sphere.

[0088] Calculate ID both The number of samples in :

[0089] t both =length(Id both )

[0090] The particle-sphere correlation matrix P between views (m,n) Elements in satisfy:

[0091]

[0092] Where: m and n are view numbers, t i is a ball in view m, t j is a ball in view n, t both is the number of samples in the intersection of two spheres, and τ is a threshold parameter, which indicates the minimum proportion of common samples required for two cross-view spheres.

[0093] Matrix Set Reflects whether any two spheres in the view overlap. The matrix set Indicates whether any two spheres between views have enough intersection. The goal is that these related sphere pairs are as close as possible in the latent space, while unrelated sphere pairs should be far apart. The sphere center is used to represent the entire sphere during the calculation process. For ease of calculation, for any two views m and n, the combined center matrix is ​​defined as:

[0094]

[0095] Then add {A m ,A n} and {P (m,n) ,P (m,n)} are concatenated into a unified mask matrix:

[0096]

[0097] Where: P (n,m) P (m,n) The transpose of .

[0098] The unified mask matrix M ensures that particles within the same view and across different views are properly considered during the contrastive learning process.

[0099] Step 3: Train the multi-view comparative clustering model to obtain a trained multi-view comparative clustering model;

[0100]

[0101] Where: Ω i represents the set of ball numbers associated with the i-th ball, φ irepresents the set of ball numbers associated with the i-th ball, M ij Indicates whether there is a correlation between the i-th ball and the j-th ball, M iz Indicates whether the ith ball is associated with the zth ball; cos(·,·) is the cosine similarity between the two vectors.

[0102] The same calculation process is performed between any two views, and the average of all losses is used as the final contrast loss. The contrast loss function is as follows:

[0103]

[0104] Where: V is the number of views, L (m,n) is the loss between views m and n.

[0105] The encoding loss is:

[0106]

[0107] Where: E v (·;θ v ) is the encoder, D v (·;φ v ) is a decoder, is, v is the parameter, φ v is a parameter, N is the number of samples in the view, V is the number of views, v is the view number, and i is the sample number.

[0108] Combining the above two loss functions with the regularization parameter λ, we get the total loss function and any gradient-based optimization method such as the Adam algorithm can be used to represent this objective function.

[0109] The loss functions include contrast loss and encoding loss;

[0110] L=L con +λL rec

[0111] Where: L is the loss function, L con is the contrast loss, L rec is the encoding loss and λ is a parameter.

[0112] By optimizing the objective function, i.e., the loss function, the multi-view features gradually have clear cluster boundaries in the latent space and show good discriminability.

[0113] The output module is used to output clustering results based on fusion features;

[0114] Feature Matrix The fusion matrix is ​​obtained by contrastive learning The k-means algorithm is directly executed on the fusion matrix to obtain the clustering label of each sample, that is, the clustering result.

[0115] Step 4: Obtain the desired clustering results based on the trained multi-view comparative clustering model.

[0116] Example

[0117] To illustrate the effectiveness of the proposed method, seven multi-view benchmark datasets are used.

[0118] BBCSport includes 544 sports news articles in 5 subject areas, with 3183-dimensional MTX features and 3203-dimensional TERMS features, forming 2 views. Caltech101-20 contains a total of 101 classes, 20 widely used 2-view classes and 2386 samples of the class are selected.

[0119] Cora contains 4 views extracted from documents: content, inbound, outbound, and reference. Scene15 consists of 4568 natural scenes divided into 15 groups. Each scene is described by three types of features: GIST, SIFT, and LBP. MNIST-USPS is a popular handwritten digit dataset containing 5000 samples with two different types of digit images. ALOI-100 consists of 10800 object images, each of which is described by 4 different features. NoisyMNIST uses the original image as the view. Figure 1 , randomly select the intra-class image of Gaussian white noise as the visual Figure 2 The specific data set is shown in Table 1.

[0120] Table 1. Multi-view datasets

[0121]

[0122] The method of the present invention is compared with the existing method, which includes

[0123] Completer (Lin et al. 2021), MFLVC (Xu et al. 2022), DealMVC (Yang et al. 2023b), DMCE (Zhao, Yang, and Nie 2023), CSPAN (Jin et al. 2023b), ADPAC (Xu et al. 2023b), SURE (Yang et al. 2023a).

[0124] Commonly used indicators are used to evaluate the results: clustering accuracy ACC, normalized mutual information NMI and purity PUR.

[0125] For each view, the encoder consists of several linear layers with ReLU activation functions between each pair of layers. Except for the BBCSport and Cora datasets, all other datasets use the same encoder structure with the dimension set to {d v ,2000,500,500,d}, where d v is the input feature dimension for each view. d is the projected feature dimension, which is the same for all views. After encoding, the input is normalized. The decoder mirrors the encoder structure. For the Cora dataset, the same dimension is used but without activation functions between layers, resulting in linear projections. For BBC Sport, given its sample size of 544, we use an encoder dimension set to {d v ,d}. In the clustering stage, the projection features of each view are fused with equal weights. Then the k-means algorithm is applied to obtain cluster labels.

[0126] The implementation of the proposed method was performed using PyTorch 2.3 (Paszke et al. 2019) on a Windows 10 operating system powered by an NVIDIA GeForce GTX 1660 Ti GPU. The Adam optimizer with a learning rate of 0.0001 and a weight decay of 0 was used. The batch size was typically set to 256 or 1024, depending on the dataset size. The regularization parameter λ is typically set to 1 in most datasets, except for BCSport and Cora, where it is adjusted to 0 due to differences in projection methods (e.g., linear or nonlinear). The threshold parameter τ is uniformly set to 0.1 for all datasets. The granularity parameter p significantly affects the experimental results and will be analyzed later. It should be emphasized that the granularity parameter reflects the average granularity, and the granular ball set may contain balls smaller than or equal to this granularity.

[0127] The results are shown in Table 2, where bold indicates the best result and underline indicates the second best result.

[0128] Table 2. Clustering results of different methods on 7 datasets

[0129]

[0130] As can be seen from Table 2, in most cases, the proposed method achieves the best or second best results in the three given indicators. Even on the NoisyMNIST dataset, the proposed method ranks about third, with a very small gap with the top two methods. Taking the Cora dataset as an example, the proposed method achieves an accuracy of 65.44%, which greatly exceeds the best comparison result of 49.07%. This proves the effectiveness and competitiveness of the proposed method.

[0131] Compared with classic multi-view contrastive learning methods (such as Completer, DealMVC, SURE), the proposed method consistently achieves more favorable clustering results on most datasets. As a representative, SURE is an instance-level contrastive learning method that focuses on the false negative problem and achieves the best or second best results on Scene-15, MNIST-USPS, ALOI-100, and NoisyMNIST datasets. It is worth noting that the proposed method does not show significant performance degradation on these four datasets and significantly outperforms other methods. In addition, on the remaining three datasets, the proposed method significantly outperforms SURE, highlighting the effectiveness of contrastive learning at the granular level.

[0132] The clustering results visualized using t-SNE on the MNIST-USPS dataset are as follows: Figure 2 As shown, it can be seen that with the increase of optimization rounds, the clustering structure becomes clearer, which shows that the method of the present invention can effectively reveal the underlying cluster structure.

[0133] To further verify the effectiveness of the proposed method at the granular level, ablation experiments were performed on the Caltech101-20 dataset. The feature dimension d was set to 128. Based on this, we adopted three experimental settings. The first setting trained the model based on the reconstruction loss only. The second setting included the reconstruction loss and the instance-level contrast loss (i.e., p = 1). The third setting included the reconstruction loss and the granular contrast loss, and the granularity parameter p was set to 2 (the proposed method).

[0134] The experimental results are shown in Table 3, from which it can be seen that the instance-level comparison method performs poorly on this dataset, while the particle-ball comparison method achieves significant improvements, indicating the feasibility and effectiveness of this method.

[0135] Table 3. Ablation experiment results on the Caltech101-20 dataset

[0136]

[0137] There are two important hyperparameters in the proposed method, the granularity parameter p and the projection feature dimension d; the former essentially reflects the average size of the granules, and when p is set to 1, it is equivalent to instance-level contrastive learning. The latter affects the amount of original feature information contained in the latent representation. If d is too small, important information may be lost, and if it is too large, it will increase the complexity of optimization and memory requirements.

[0138] Experiments were performed on Caltech101-20 and Cora with varying parameter p in the range [1, 2, 3, 4, 8, 16] and parameter d in the range [8, 16, 32, 64, 128, 256]. The parameters for all datasets were selected from these ranges. The results are shown in Figure 3 As shown in the figure, it can be seen that the method of the present invention performs well when p is set to 2. As p increases, the performance gradually decreases because the larger particle balls are no longer effectively associated based only on the overlap and intersection size. When p becomes too large, the method of the present invention essentially degenerates into a cluster-level contrast method, where reducing cluster assignment differences may be better.

[0139] So the parameter p is usually set to 1, 2 or 4. The parameter d does not have a big impact on the result and is usually set to 64.

[0140] The convergence of the proposed method on the Cora dataset is evaluated by tracking the loss value and the corresponding clustering performance as the number of epochs increases. The total loss value gradually decreases and converges within 100 epochs. These results demonstrate the strong convergence performance of the proposed method.

[0141] The present invention performs contrastive learning at the granule level, avoids directly using adjacent samples to construct negative pairs, and retains the local structural information of the sample set, overcoming the shortcomings of instance-level and cluster-level methods. The granule construction method of the present invention is different from the classical method of continuously dividing the data set until the minimum granularity is reached. The sample set can be directly divided into multiple granules according to the granularity parameter, avoiding the disadvantage that non-adjacent samples are divided into the same granule in the boundary area.

[0142] The present invention uses spheres to model the local structure of the sample set, and establishes intra-view and inter-view sphere connections based on the overlap and intersection sizes, respectively. By bringing the connected spheres closer in the latent space, the proposed model learns highly discriminative features.

Claims

1. A multi-view contrast clustering method based on granular sphere, characterized in that: The following steps are involved: Step 1: Obtain multi-view data to form a data set; Step 2: construct a multi-view comparison clustering model, which includes a data processing module, a particle sphere generation module and an output module; The data processing module is used to reduce the dimension of the multi-view data in the data set to obtain the multi-view data features; The particle sphere generation module is used to construct a particle sphere set according to each view corresponding to the multi-view data features, and to construct a particle sphere overlap relationship matrix within each view and a particle sphere correlation matrix between views; The mask matrix is ​​obtained according to the particle-ball overlap relationship matrix within the view and the particle-ball correlation matrix between views, and the fusion feature is obtained through contrast learning; The output module is used to output clustering results based on fusion features; Step 3: Train the multi-view comparative clustering model to obtain a trained multi-view comparative clustering model; Step 4: Obtain the desired clustering results based on the trained multi-view comparative clustering model.

2. The method of majority graph comparison clustering based on particle sphere according to claim 1, characterized in that: The process of constructing the particle sphere set in step 2 is as follows: In each view, k clusters, i.e., k spheres, are obtained by the k-means clustering method. The k spheres constitute a sphere set. The above process is repeated for each view to obtain the entire sphere set.

3. The multi-view contrast clustering method based on particle sphere according to claim 2, characterized in that: The center c of the sphere i and radius r i The calculation method is as follows: Where: i is the number of the ball, n i is the number of samples in the sphere, j is the serial number of the samples contained in the sphere, x j is the jth sample of the current ball.

4. The multi-view contrast clustering method based on particle sphere according to claim 3, characterized in that: The overlapping relationship matrix A v Elements in The following relations are satisfied: Where: v is the view number; The distance between the particles is calculated as follows: Where: is the distance between the i-th sphere and the j-th sphere in the v-th view, is the center of the i-th sphere in the v-th view, is the center of the jth sphere in the vth view.

5. The multi-view contrast clustering method based on particle sphere according to claim 4, characterized in that: The inter-view particle-sphere correlation matrix P (m,n) Elements in satisfy: Where: m and n are view numbers, t i is a ball in view m, t j is a ball in view n, t both is the number of samples in the intersection of two spheres, and τ is the threshold parameter; in: t both =length(Id both ) Where: Id both is the common sample set of the i-th sphere in view m and the j-th sphere in view n.

6. The multi-view contrast clustering method based on particle sphere according to claim 5, characterized in that: The mask matrix M is: Where: P (n,m) P (m,n) The transpose of .

7. The multi-view contrast clustering method based on particle sphere according to claim 1, characterized in that: The loss function in the training process includes contrast loss and encoding loss; L=L con +λL rec Where: L is the loss function, L con is the contrast loss, L rec is the encoding loss and λ is the balancing hyperparameter.

8. The multi-view contrast clustering method based on particle sphere according to claim 7, characterized in that: The contrast loss function is as follows: Where: V is the number of views, L (m,n) is the loss between views m and n.

9. The multi-view contrast clustering method based on particle sphere according to claim 7, characterized in that: The coding loss is: Where: E v (·;θ v ) is the encoder, D v (·;φ v ) is a decoder, is, θv is a parameter, φ v is a parameter, N is the number of samples in the view, V is the number of views, v is the view number, and i is the sample number.

10. A system using the multi-view comparison clustering method based on particle spheres as claimed in any one of claims 1 to 9, characterized in that: It includes data acquisition and processing module, pellet generation module and output module; The data acquisition and processing module is used to acquire a multi-view data set, perform dimensionality reduction on the multi-view data in the data set, and obtain multi-view data features; The particle sphere generation module is used to construct a particle sphere set according to each view corresponding to the multi-view data features, and to construct a particle sphere overlap relationship matrix within each view and a particle sphere correlation matrix between views; The mask matrix is ​​obtained according to the particle-ball overlap relationship matrix within the view and the particle-ball correlation matrix between views, and the fusion feature is obtained through contrast learning; The output module is used to output clustering results based on fusion features.

Images