Plug-in cross-view pre-learning method for multi-view clustering

Abstract

The invention belongs to the technical field of data analysis and application, and particularly relates to a multi-view clustering-oriented plug-in cross-view pre-learning method, which comprises the following steps of: firstly, obtaining coded representation of original input data on each view by adopting a multi-view automatic encoder; then, complementary consistency information among multiple views is introduced in a cross-view comparative learning mode; and finally, performing model supervision by taking multi-view reconstruction loss and cross-view contrast learning loss as total task loss through joint optimization. According to the method, by introducing cross-view comparative learning and an insertion type pre-learning method, multi-view information is fully considered, and the multi-view information is comprehensively and effectively integrated in a feature learning stage, so that the expression ability and clustering performance of the model are improved, the information loss phenomenon is reduced, and a more accurate and stable multi-view clustering result is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of data analysis and application, and in particular relates to a plug-in cross-view pre-learning method for multi-view clustering. Background Art

[0002] In recent years, multi-view clustering methods, as an integral part of data analysis, have effectively supported data analysis and applications in many fields such as text and images. The core idea is to integrate and utilize data from multiple perspectives of the data source, reveal the complex relationship between the data, mine more abundant and accurate information, and use this information for effective clustering analysis. With the development of deep learning, multi-view deep clustering methods have gradually become a research hotspot.

[0003] Multi-view deep clustering methods are mainly divided into two stages: representation learning and clustering tasks. Among them, the representation learning stage can be divided into two key steps: representation pre-learning and representation fine-tuning. Representation pre-learning aims to optimize the representation of raw data, remove noise and capture key information through techniques such as dimensionality reduction, feature selection and transformation. Representation fine-tuning further adjusts the representation according to the specific clustering task, including adjusting feature weights and enhancing the expressiveness of specific features. In the clustering task stage, the data is grouped into categories or clusters with similar characteristics for further data analysis and understanding. A large number of recent works have shown that the improvement in the performance of sample representation has greatly improved the accuracy of clustering division, making the multi-view deep clustering method have better performance results.

[0004] For real data, there is a mutually complementary and reinforcing relationship between multiple views of the data. For example, for text data, the author view and the content view complement and enhance each other, providing different but related information. The author view focuses on the creator or speaker of the text, including his or her background, position, and intention, while the content view focuses on the specific information, themes, and arguments conveyed by the text. The author view explains the background and emotional tendency of the text, and the content view provides more specific semantic information and thematic content. Combining the two can more comprehensively understand the text, evaluate its credibility, perform sentiment analysis, and realize applications such as content recommendation. However, the multi-view information learning of existing methods is mostly focused on the representation fine-tuning part. In the representation pre-learning process, most of them are independently learned from the original data of each view, ignoring the cross-view complementary and reinforcing relationship between multi-view information. This independent pre-learning method makes the model unable to fully utilize the rich information from different views, resulting in insufficient utilization of semantic information, which in turn affects clustering performance.

[0005] Among the existing methods, there are multi-view clustering methods based on traditional pre-learning methods, such as the document "XIE Shengli, CHEN Hongda, GAO Junli, et al. Distribution-Aligned Variational Autoencoder[J]. Journal of Computer Science and Technology, 2023, 46 (5): 5.", in which the distribution-aligned variational autoencoder is used to replace the traditional autoencoder in the representation pre-learning process, so that the feature learning process is closer to reality and the text features are described in the form of probability distribution. This method improves the learning effect of the model to a certain extent, making the learned features more expressive and robust. Another example is the literature "LIU Jianing, ZENG Jingxia. Robust orthogonally constrained multi-view subspace clustering algorithm [J]. Computer System Applications, 2024, 33(04): 171-178. DOI:10.15888 / j.cnki.cs a.009448. (LIU Jianing, ZENG Jingxia. Robust orthogonally constrained multi-view subspace clustering algorithm [J]. Computer System Applications, 2024, 33(04): 171-178. DOI:10.15888 / j.cnki. csa.009448. In the method disclosed in this document, a robust representation is obtained by eigendecomposition in the representation pre-learning process; then a joint latent subspace representation is constructed for multiple views, spectral rotation is used to obtain clustering results, and orthogonal constraints are applied to the partition matrix to reconstruct the subspace, thereby improving clustering performance. The methods disclosed in the above two documents do not consider multi-view information in the pre-learning process, but focus on feature learning in a single view.

[0006] Among the existing methods, the multi-view clustering method based on contrastive learning, such as the document [5]Bai, R., Huang, R., Qin, Y., Chen, Y. (2022). Multi-view Document Clustering with JointContrastive Learning. In: Lu, W., Huang, S., Hong, Y., Zhou, X. (eds) NaturalLanguage Processing and Chinese Computing. NLPCC 2022. Lecture Notes inComputer Science(), vol 13551. Springer, Cham. https: / / doi.org / 10. 1007 / 978-3-031-17120-8_55, proposed the MCJC model, which uses joint contrastive learning in the representation fine-tuning stage and introduces feature contrast loss and clustering contrast loss for joint optimization. Although it considers multi-view information in representation learning, it is only in the representation fine-tuning part mentioned above, and the semantic information is not fully utilized. For example, the paper "P. Hu, D. Peng, Y. Sang, and Y. Xiang, "Incomplete Multi-view Clustering via Prototype-basedImputation," arXiv preprint arXiv: 2301.11045v2, Jan .2023." proposes the ProImp model, which uses dual attention layers and dual contrastive learning losses. When a view is lost, it can use representative samples in the missing view and the sample-representative sample relationship inherited from the observed view to perform data recovery. The methods disclosed in the above two papers only introduce contrastive learning in the representation fine-tuning stage, but ignore the cross-view complementary enhancement relationship between multi-view information in the representation pre-learning stage, resulting in incomplete feature learning.

[0007] In summary, the existing methods have the following technical problems:

[0008] (1) Information loss:

[0009] In multi-view clustering, the existing methods only extract features from a single view and cannot effectively capture the relationship and information between different views. The lack of multi-view information integration results in the inability to fully reflect the diversity and complexity of the data. Therefore, the existing methods often encounter the problem of information loss when performing multi-view clustering, that is, they fail to fully consider multi-view information in the feature learning stage, resulting in some data information being ignored or lost.

[0010] (2) Clustering performance decreases:

[0011] When performing multi-view clustering, the existing technology does not fully consider multi-view information in the feature learning stage, which limits the effect of feature learning, thereby affecting the execution efficiency and accuracy of subsequent clustering tasks, resulting in a decrease in clustering performance and inaccurate and unstable clustering results.

[0012] (3) Insufficient model expression ability:

[0013] When performing multi-view clustering, the existing technology only conducts feature learning from the perspective of a single view and fails to comprehensively consider the information of multiple views, which limits the expressive power of the model and makes it unable to capture the diversity and complexity of the data well. Summary of the invention

[0014] The present invention aims to solve the problems of information loss, reduced clustering performance and insufficient model expression in existing methods, and provides a plug-in cross-view pre-learning method for multi-view clustering. The present invention introduces cross-view contrastive learning and plug-in pre-learning methods, fully considers multi-view information, and comprehensively and effectively integrates multi-view information in the feature learning stage, thereby improving the expression ability and clustering performance of the model, reducing information loss, and achieving more accurate and stable multi-view clustering results.

[0015] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0016] The present invention provides a plug-in cross-view pre-learning method for multi-view clustering, which mainly includes the following steps:

[0017] Firstly, a multi-view autoencoder is used to obtain the encoded representation of the original input data on each view. Then, a cross-view contrastive learning method is used to introduce the complementary consistency information between multiple views. Finally, the multi-view reconstruction loss and the cross-view contrastive learning loss are jointly optimized as the total task loss for model supervision.

[0018] Furthermore, the multi-view autoencoder is a denoising autoencoder, a convolutional autoencoder, a long short-term memory autoencoder or an adversarial autoencoder.

[0019] Furthermore, in the multi-view clustering task, given a Views and A multi-view dataset of samples, and Representation View The original input data, Representation View The input dimensions of the original input data; for a given view The original input data , the mathematical expression of the learning representation of the corresponding encoding layer is:

[0020] (1) ,

[0021] in, Representation View Encoder, Representation View The parameters corresponding to the encoder, Representation View The original input data, Representation View The encoded representation of the original input data.

[0022] Furthermore, for a given view And its corresponding encoding representation , the mathematical expression of the specific decoding process is:

[0023] (2) ,

[0024] in, Represents a decoder, Indicates the parameters corresponding to the decoder, Representation View Reconstructed data; for a given view , and its view reconstruction self-supervision loss function is defined as shown in formula (3):

[0025] (3) ,

[0026] in, represents the L2 norm formula, N represents the number of samples;

[0027] For a given For a multi-view dataset with multiple views, the multi-view reconstruction loss is the sum of the self-supervised reconstruction losses of all views, that is:

[0028] (4).

[0029] Furthermore, when using a cross-view contrastive learning approach to introduce complementary consistency information between multiple views, we first use the intra-view reconstruction of each view to retain the feature learning within the view, and then use the complementary consistency information between multiple views to construct a cross-view contrastive learning loss to promote multi-view representation learning.

[0030] Furthermore, when using cross-view contrastive learning to introduce complementary consistency information between multiple views, the cluster-oriented information is regarded as a representation of different cluster distributions on each sample, and a two-layer nonlinear fully connected layer is used. From the view The encoding representation of the original input data set Extract the cluster distribution representation of samples from , the process is formalized as:

[0031] (5) ,

[0032] in, represents the number of clusters in the dataset, No. Column Representation View The The cluster distribution is represented by ;

[0033] The loss function of formula (6) is used to learn the consistency of the cluster distribution representation between multiple views:

[0034] l c ( k , v ) =− log exp( sim ( c k v , c k v ¯ ) / τ c ) ∑ l = 1 K [exp( sim ( c k v , c l v ) / τ c ) + exp( sim ( c k v , c l v ¯ ) / τ c )] (6) ,

[0035] in, Representation temperature parameter, used to control the softness of cluster distribution representation between multiple views; Represents the cosine similarity of the positive example pair; the calculation of the negative example pair includes two cases, one is represents the cosine similarity of negative example pairs in the same view, and the second is Represents the cosine similarity of negative example pairs between different views;

[0036] Consider the entropy of cluster assignment probability to avoid that most of the clustering results are assigned to the same cluster. The corresponding calculation formula is as follows:

[0037] (7)

[0038] in, The calculation formula representing the probability of cluster distribution;

[0039] The specific calculation formula is as follows:

[0040] (8)

[0041] in, represents the kth cluster distribution of the vth view and the j'th sample, represents the overall clustering distribution of the v-th view.

[0042] Therefore, the mathematical expression of cross-view contrastive learning loss is as follows:

[0043] (9)

[0044] in, and Respectively represent the Figure 1 And Vision Figure 2 The original data.

[0045] Furthermore, in the process of jointly optimizing the multi-view reconstruction loss and the cross-view contrastive learning loss as the total task loss for model supervision training, on the basis of relatively independent view reconstruction, the consistency between views is further considered, the corresponding cross-view contrastive supervision target is designed, and the model parameters are optimized by minimizing the joint loss of the two; the mathematical expression of the corresponding total task loss is:

[0046] (10)

[0047] in, represents the multi-view reconstruction loss, represents the cross-view contrastive learning loss.

[0048] The beneficial effects of the present invention are:

[0049] In view of the fact that existing work cannot fully utilize and integrate multi-view information in the pre-learning process, and ignores the problem that the cross-view complementary enhancement relationship between multi-view information will affect the clustering performance in the representation pre-learning stage, the present invention proposes a plug-in cross-view pre-learning method for multi-view clustering, which fully considers the cross-view complementary enhancement relationship between multi-view information in the representation pre-learning stage to improve the representation learning ability of the model and thus optimize the clustering performance. Due to the excellent performance of contrastive learning in learning the consistency of view representation, the present invention introduces the consistency information between multiple views through cross-view contrastive learning in the representation pre-learning stage, that is, more semantic information is considered in the process of multi-view representation pre-learning.

[0050] In addition, in order to better supervise the learning of representations, a cross-view supervision module is designed. This cross-view supervision module realizes the fusion of view reconstruction self-loss and cross-view contrast loss, and considers cross-view information while supervising view self-feature learning, providing clearer task guidance for the model, so that the guidance task of feature learning is transformed from a single reconstruction to a more comprehensive consideration of the semantic information of multiple views, ensuring more accuracy and pertinence when learning multi-view information. At the same time, the plug-in cross-view pre-learning method for multi-view clustering proposed in the present invention has high scalability and can be used in combination with other clustering algorithms in a plug-and-play manner, providing flexible and effective solutions for different application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A flow chart of a plug-in cross-view pre-learning method for multi-view clustering provided by the present invention.

[0052] Figure 2 The NMI performance of MCJC in each data set.

[0053] Figure 3 The CA performance of MCJC on each dataset. DETAILED DESCRIPTION

[0054] The present invention provides a plug-in cross-view pre-learning method for multi-view clustering. Through the plug-in pre-learning method, before the clustering task is executed, that is, the pre-learning part of the data, through the two parts of the view reconstruction self-supervision module and the cross-view comparison supervision module, contrast learning is used to obtain representation learning that introduces multi-view information, and features that can extract information from the data itself and introduce multi-view information are learned to improve the performance of downstream tasks, thereby obtaining better clustering effects under various clustering methods.

[0055] The present invention provides a plug-in cross-view pre-learning method for multi-view clustering, which mainly includes two parts: a multi-view feature learning module and a cross-view supervision module.

[0056] Among them, the multi-view feature learning module is mainly used to use the multi-view autoencoder to obtain the encoded representation, i.e., feature representation, of the original input data on each view; the cross-view supervision module is mainly used to introduce the complementary consistency information between multiple views by contrastive learning, and jointly optimizes the multi-view reconstruction loss and the cross-view contrastive learning loss as the total task loss for model supervision.

[0057] Among them, the specific functions and effects of the multi-view feature learning module are as follows:

[0058] For different types of data, different unsupervised learning methods can be used, such as denoising autoencoders, convolutional autoencoders, long short-term memory (LSTM) autoencoders, and adversarial autoencoders. In order to adapt to different data types, the present invention uses a basic multi-view autoencoder to learn the representation of raw data.

[0059] In the multi-view clustering task, given a Views and multi-view dataset of samples, and { } represents the view The original input data is Representation View The input dimensions of the original input data.

[0060] For a given view The original input data , the mathematical expression of the learning representation of the corresponding encoding layer is:

[0061] (1) ,

[0062] in, Representation View Encoder, Representation View The parameters corresponding to the encoder, Representation View The original input data, Representation View The encoded representation of the original input data.

[0063] Among them, the specific functions and roles of the cross-view supervision module are as follows:

[0064] The cross-view supervision module aims to use the complementary consistency information between multiple views in a contrastive learning manner to learn more discriminative view representations. The cross-view supervision module consists of two parts: the view reconstruction self-supervision module and the cross-view contrast supervision module. The view reconstruction self-supervision module uses the intra-view reconstruction of each view to retain the feature learning inside the view; the cross-view contrast supervision module uses the complementary consistency information between multiple views to construct a cross-view contrast learning loss to promote multi-view representation learning.

[0065] Among them, the specific operation process of the view reconstruction self-supervision module is as follows:

[0066] The view reconstruction self-supervision module is mainly composed of a decoder corresponding to the multi-view autoencoder. Its network structure is similar to that of the multi-view autoencoder.

[0067] For a given view And its corresponding encoding representation , the mathematical expression of the specific decoding process is:

[0068] (2) ,

[0069] in, Represents a decoder, Indicates the parameters corresponding to the decoder, Representation View By reducing the view Reconstructed data With View The original input data The gap between the two is to learn the neural network layer parameters that best express the data features and obtain the encoding representation of the optimal view. . Therefore, for a given view , and its view reconstruction self-supervision loss function is defined as shown in formula (3):

[0070] (3)

[0071] in, represents the L2 norm formula, and N represents the number of samples.

[0072] Thus, for a given For a multi-view dataset with multiple views, the multi-view reconstruction loss is the sum of the self-supervised reconstruction losses of all views, that is:

[0073] (4)

[0074] Among them, the specific operation process of the cross-view comparison supervision module is as follows:

[0075] The cross-view contrast supervision module is used to introduce complementary consistency information between multiple views. For traditional multi-view data, a graph model can be used, in which nodes represent text samples and edges represent the relationship between samples. Multimodal fusion can also be performed, including serial, parallel, attention mechanisms, etc., by using multimodal fusion mechanisms to integrate information between different views. However, the above methods have many prerequisites and are not suitable for grafting into lightweight pre-learning models. Therefore, the present invention adopts a cross-view contrast learning method to introduce complementary consistency information between multiple views, that is, more semantic information is considered in the process of multi-view representation pre-learning. Inspired by the "label is also a representation", the present invention regards cluster-oriented information as a representation of different cluster distributions on each sample. Specifically, the present invention uses two layers of nonlinear fully connected layers. From the view The encoding representation of the original input data set Extract the cluster distribution representation of samples from . For a given view , the process can be formalized as:

[0076] (5)

[0077] in, represents the number of clusters in the dataset, No. Column Representation View The The cluster distribution is represented by .

[0078] The present invention uses the following loss function to learn the consistency of the cluster distribution representation between the maximized multiple views:

[0079] l c ( k , v ) =− log exp( sim ( c k v , c k v ¯ ) / τ c ) ∑ l = 1 K [exp( sim ( c k v , c l v ) / τ c ) + exp( sim ( c k v , c l v ¯ ) / τ c )] (6)

[0080] in, It represents the temperature parameter, which is used to control the softness of the representation of cluster distribution between multiple views. It is fixed to 1.0 in the experiment. represents the cosine similarity of the positive example pair, which is formed by the cluster distribution representation of different views from the same cluster. The calculation of the negative example pair includes two cases, one is represents the cosine similarity of negative example pairs in the same view, and the second is In order to avoid that most of the clustering results are assigned to the same cluster, the present invention also considers the entropy of cluster assignment probability, and the corresponding calculation formula is as follows:

[0081] (7)

[0082] in, The calculation formula representing the probability of cluster distribution;

[0083] The specific calculation formula is as follows:

[0084] (8)

[0085] in, represents the kth cluster distribution of the vth view and the j'th sample, represents the overall clustering distribution of the v-th view.

[0086] In summary, the mathematical expression of cross-view contrastive learning loss is as follows:

[0087] (9)

[0088] in, and Respectively represent the Figure 1 And Vision Figure 2 The original data.

[0089] Then, a joint optimization is performed to use the multi-view reconstruction loss and the cross-view contrastive learning loss as the total task loss for model supervision. The specific operation process is as follows:

[0090] During the training process, different from the traditional reconstruction loss, this paper adopts a joint optimization strategy to combine the multi-view reconstruction loss and the cross-view contrast learning loss as the task goal of the model in the representation pre-learning stage. Specifically, on the basis of relatively independent view reconstruction, the consistency between views is further considered, the corresponding cross-view contrast supervision goal is designed, and the model parameters are optimized by minimizing the joint loss of the two to achieve better clustering effect. Among them, the mathematical expression of the corresponding total task loss is:

[0091] (10)

[0092] in, represents the multi-view reconstruction loss, Represents the cross-view contrastive learning loss. Through the joint optimization of the two, the supervision information within and between views in the multi-view data can be comprehensively considered during the training process, effectively improving the model's expressiveness in multi-view data and guiding the model to learn the potential structure of the data, thereby better improving the clustering performance.

[0093] like Figure 1 As shown, 2 Indicates that the total loss is the weighted sum of the two losses, and the weight distribution can be manually specified. The present invention takes the data set of two views as an example, and Respectively represent the Figure 1 And Vision Figure 2 The original data, and Represents the original data and The encoding representation is the feature representation. and Respectively represent the Figure 1 And Vision Figure 2 The reconstructed data of the image is used to introduce in detail a plug-in cross-view pre-learning method for multi-view clustering of the present invention. Specifically: for a given view Figure 1 And Vision Figure 2 The original data and After being encoded by the encoder, the corresponding encoding representation is obtained and , and then the encoding representation and After being decoded by the decoder, the video Figure 1 And Vision Figure 2 Reconstructed data and ; Use the view reconstruction self-supervision module to calculate the multi-view reconstruction loss; use the cross-view contrast supervision module to calculate the cross-view contrast learning loss; then perform joint optimization to use the multi-view reconstruction loss and the cross-view contrast learning loss as the total task loss for model supervision.

[0094] The present invention selects the MCJC benchmark model as the basis of the experiment, and designs comparative experiments on various data sets under the same environment and parameter configuration. Through careful comparative analysis, it is found that the pre-learning representation has a huge impact on the clustering performance. Sufficient pre-learning representation enables the model to show stable and efficient results in both NMI and CA key performance indicators. To further verify the effect of the pre-learning representation stage, the present invention performs the pre-learning representation stage for each data set with different numbers of iterations. The results are shown in Figure 2. Figure 2 and Figure 3 As shown in Figure 1, experimental results on five different datasets (Movies, UCI, AMiner-700, AMiner, Reuters) consistently show that the fully pre-learned model has significant improvements in both clustering accuracy and normalized mutual information. The performance improvement on the UCI and AMiner-700 datasets is particularly significant, which further confirms the importance of sufficient and in-depth pre-learning representations when dealing with datasets with high diversity and complexity.

[0095] Table 1 presents the clustering results of the plug-in cross-view pre-learning method for multi-view clustering proposed in the present invention compared with the prior art on various real data sets. PCRL_DEMVC, PCRL_MCJC, PCRL_CVCL and PCRL_ProImp demonstrate the effect of applying the plug-in cross-view pre-learning method for multi-view clustering proposed in the present invention to different existing models. The results show that from Movies, UCI, AMiner-700, AMiner to Reuters datasets, the plug-in cross-view pre-learning method for multi-view clustering proposed in the present invention outperforms the MCJC benchmark model in two core indicators, normalized mutual information (NMI) and unsupervised clustering accuracy (CA), highlighting its significant advantage in improving clustering performance. The reason for achieving this effect is that: in the pre-learning stage, a plug-in cross-view pre-learning method for multi-view clustering proposed in the present invention is used to realize cross-view comparative learning and integrate multi-view information. This approach can better capture the diversity of data, make full use of the rich information in each view, provide clear task guidance for representation learning, and ensure the accuracy and pertinence of the model when processing multi-view information.

[0096] Table 1 Comparison of clustering results (%)

[0097]

[0098] Under the inventive principle of this pre-learning process, cross-view contrast information can be introduced in the process of feature learning to obtain more complete semantic learning and better clustering effect. On this basis, technicians in this field can use different similarity metrics or modify the formula of contrast loss to obtain better cross-view contrast learning effect. The scope of contrast learning is very wide and is not limited to the contrast loss method given in this patent. In addition, other technicians can also make modifications in the downstream clustering task process and combine our pre-learning model to obtain better clustering effect.

[0099] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A plug-in cross-view pre-learning method for multi-view clustering, characterized in that: The following steps are involved: Firstly, a multi-view autoencoder is used to obtain the encoded representation of the original input data on each view. Then, a cross-view contrastive learning method is used to introduce the complementary consistency information between multiple views. Finally, the multi-view reconstruction loss and the cross-view contrastive learning loss are jointly optimized as the total task loss for model supervision.

2. The plug-in cross-view pre-learning method for multi-view clustering according to claim 1, characterized in that: The multi-view autoencoder is a denoising autoencoder, a convolutional autoencoder, a long short-term memory autoencoder or an adversarial autoencoder.

3. The plug-in cross-view pre-learning method for multi-view clustering according to claim 1, characterized in that: In the multi-view clustering task, given a Views and multi-view dataset of samples, and { } represents the view The original input data, Representation View The input dimensions of the original input data; for a given view The original input data , the mathematical expression of the learning representation of the corresponding encoding layer is: (1)， in, Representation View Encoder, Representation View The parameters corresponding to the encoder, Representation View The original input data, Representation View The encoded representation of the original input data.

4. The plug-in cross-view pre-learning method for multi-view clustering according to claim 3, characterized in that: For a given view And its corresponding encoding representation , the mathematical expression of the specific decoding process is: (2)， in, Represents a decoder, Indicates the parameters corresponding to the decoder, Representation View Reconstructed data; for a given view , and its view reconstruction self-supervision loss function is defined as shown in formula (3): (3)， in, represents the L2 norm formula, N represents the number of samples; For a given For a multi-view dataset with multiple views, the multi-view reconstruction loss is the sum of the self-supervised reconstruction losses of all views, that is: (4)。 5. The plug-in cross-view pre-learning method for multi-view clustering according to claim 1, characterized in that: When using cross-view contrastive learning to introduce complementary consistency information between multiple views, we first use the intra-view reconstruction of each view to retain the feature learning within the view, and then use the complementary consistency information between multiple views to construct a cross-view contrastive learning loss to promote multi-view representation learning.

6. The plug-in cross-view pre-learning method for multi-view clustering according to claim 1, characterized in that: When using cross-view contrastive learning to introduce complementary consistency information between multiple views, the cluster-oriented information is regarded as a representation of different cluster distributions on each sample, and a two-layer nonlinear fully connected layer is used. From the view The encoding representation of the original input data set Extract the cluster distribution representation of samples from , the process is formalized as: (5)， in, represents the number of clusters in the dataset, No. Column Representation View The The cluster distribution is represented by ; The loss function of formula (6) is used to learn the consistency of the cluster distribution representation between multiple views: (6)， in, Representation temperature parameter, used to control the softness of cluster distribution representation between multiple views; Represents the cosine similarity of the positive example pair; the calculation of the negative example pair includes two cases, one is represents the cosine similarity of negative example pairs in the same view, and the second is Represents the cosine similarity of negative example pairs between different views; Consider the entropy of cluster assignment probability to avoid that most of the clustering results are assigned to the same cluster. The corresponding calculation formula is as follows: (7)， in, The calculation formula representing the probability of cluster distribution; The specific calculation formula is as follows: (8)， in, represents the kth cluster distribution of the vth view and the j'th sample, represents the total cluster distribution of the v-th view; Therefore, the mathematical expression of cross-view contrastive learning loss is as follows: (9)， in, and Represent the original data of view 1 and view 2 respectively.

7. The plug-in cross-view pre-learning method for multi-view clustering according to claim 6, characterized in that: In the joint optimization, the multi-view reconstruction loss and the cross-view contrastive learning loss are used as the total task loss for model supervision training. On the basis of relatively independent view reconstruction, the consistency between views is further considered, the corresponding cross-view contrastive supervision target is designed, and the model parameters are optimized by minimizing the joint loss of the two. The mathematical expression of the corresponding total task loss is: (10)， in, represents the multi-view reconstruction loss, represents the cross-view contrastive learning loss.

Images