A plug-in cross-view pre-learning method for multi-view clustering

Abstract

The application belongs to the technical field of data analysis and application, and particularly relates to a plug-in cross-view pre-learning method for multi-view clustering, which comprises the following steps: firstly, a multi-view auto-encoder is used to obtain the encoding representation of original input data on each view; then, a cross-view contrast learning method is used to introduce complementary consistency information between multi-views; finally, joint optimization is performed to take the multi-view reconstruction loss and the cross-view contrast learning loss as the total task loss for model supervision. Through the introduction of the cross-view contrast learning and the plug-in pre-learning method, the application fully considers the multi-view information, effectively integrates the multi-view information in the feature learning stage, thereby improving the expression ability and clustering performance of the model, reducing the information loss phenomenon, and realizing more accurate and stable multi-view clustering results.

Description

TECHNICAL FIELD

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

[0002] In recent years, as a component of data analysis, multi-view clustering methods effectively support data analysis and application in many fields such as text and image. 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 rich and accurate information, and use these information for effective clustering analysis. With the development of deep learning, multi-view deep clustering method has gradually become a research hotspot.

[0003] Multi-view deep clustering method mainly includes two stages: representation learning and clustering task. 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 the original data by dimensionality reduction, feature selection and transformation, and remove noise and capture key information. Representation fine-tuning further adjusts the representation according to the specific clustering task, including adjusting the feature weight and enhancing the expression ability of specific features. The clustering task stage groups data into categories or clusters with similar features for further data analysis and understanding. Recent work shows that the performance of sample representation greatly improves the accuracy of clustering division, making multi-view deep clustering method have better performance results.

[0004] For real data, there is a complementary and enhanced relationship between the 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 their background, position and intention, while the content view focuses on the specific information, theme and argument 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 theme content. Integrating both can more comprehensively understand the text, evaluate its credibility, perform sentiment analysis, and achieve content recommendation and other applications. However, the multi-view information learning of existing methods is mostly placed in the representation fine-tuning part, and most of the representation pre-learning is independently learned from the original data of each view, ignoring the cross-view complementary and enhanced relationship between multi-view information. This independent pre-learning method makes the model unable to fully utilize the rich information from different views, leading to insufficient utilization of semantic information and affecting clustering performance.

[0005] In existing methods, 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 method disclosed in the document uses a distribution-aligned variational autoencoder 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 some extent, and makes the learned features more expressive and robust. For example, the document “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.”, in which the method disclosed in the document uses a feature decomposition method to obtain a robust representation in the representation pre-learning process; then a joint latent subspace representation is constructed for multiple views, a spectral rotation is used to obtain clustering results, and an orthogonal constraint is used on the partition matrix to reconstruct the subspace, thereby improving the 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 within a single view.

[0006] Among existing methods, multi-view clustering methods based on contrastive learning, such as the paper "[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 in Computer 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 contrastive loss and cluster contrastive loss for joint optimization. Although it considers multi-view information in representation learning, it only applies to 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-based Imputation,' arXiv preprint arXiv: 2301.11045v2, Jan. 2023." proposes the ProImp model, which employs dual attention layers and dual contrastive learning loss. When a view is lost, it can use representative samples from 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, while neglecting the cross-view supplementation and enhancement relationships between multi-view information in the representation pre-learning stage, leading to incomplete feature learning.

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

[0008] (1) Information loss:

[0009] In multi-view clustering, since existing methods only perform feature extraction from a single view perspective, they cannot effectively capture the relationship and information between different views, and the lack of multi-view information integration leads to the inability to fully reflect the diversity and complexity of the data. Therefore, existing methods often encounter information loss problems when performing multi-view clustering, that is, the multi-view information is not fully considered in the feature learning stage, resulting in some data information being ignored or lost.

[0010] (2) Clustering performance is reduced:

[0011] In the prior art, the multi-view information is not fully considered in the feature learning stage when performing multi-view clustering, which limits the effectiveness of feature learning and affects the execution efficiency and accuracy of subsequent clustering tasks, resulting in reduced clustering performance and inaccurate and unstable clustering results.

[0012] (3) Insufficient model expression ability:

[0013] In the prior art, the feature learning process is only performed from a single view perspective, and the information of multiple views is not comprehensively considered, which limits the expression ability of the model and makes it difficult to capture the diversity and complexity of the data. SUMMARY

[0014] The present application provides a plug-in cross-view pre-learning method for multi-view clustering to solve the problems of information loss, reduced clustering performance, and insufficient model expression ability in existing methods. By introducing cross-view contrast learning and plug-in pre-learning methods, the present application fully considers multi-view information 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 solutions adopted by the present application to solve the technical problems are as follows:

[0016] The plug-in cross-view pre-learning method for multi-view clustering provided by the present application mainly includes the following steps:

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

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

[0019] Furthermore, in multi-view clustering tasks, given a set of views... Each view and A multi-view dataset of samples, and with Represents a view The original input data, Represents a view The input dimension of the original input data; for a given view raw input data The mathematical expression for the learned representation of the corresponding encoding layer is:

[0020] (1),

[0021] in, Represents a view Encoder, Represents a view The parameters corresponding to the encoder, Represents a view The original input data, Represents a view The encoded representation of the original input data.

[0022] Furthermore, for a given view and its corresponding encoding representation The mathematical expression for its specific decoding process is as follows:

[0023] (2),

[0024] in, Indicates decoder, This indicates the parameters corresponding to the decoder. Represents a view Reconstructing data; for a given view The view reconstruction self-supervised loss function is defined as shown in equation (3):

[0025] (3),

[0026] in, This represents the L2 norm formula, where N represents the sample size.

[0027] For a given contain 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, i.e.:

[0028] (4).

[0029] Further, when introducing the complementary consistency information between multi-views by the cross-view contrast learning manner, the view-intra reconstruction of each view is first used to retain the feature learning within the view, and then the complementary consistency information between multi-views is used to construct the cross-view contrast learning loss, so as to promote the multi-view representation learning.

[0030] Further, when introducing the complementary consistency information between multi-views by the cross-view contrast learning manner, the class cluster oriented information is regarded as the different cluster distribution representation on each sample, and a two-layer nonlinear fully connected layer is used The cluster distribution representation of the sample is extracted from the encoded representation dataset of the original input data of the view For a given view , the process is formalized as:

[0031] (5),

[0032] wherein, denotes the number of class clusters in the dataset, the i-th cluster distribution representation in the view is denoted as ;

[0033] The loss function of formula (6) is used to learn the consistency of the maximized multi-view cluster distribution representation:

[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] wherein, denotes a temperature parameter for controlling the softness of the multi-view cluster distribution representation; denotes the cosine similarity of the positive pair; the calculation of the negative pair contains two cases, one is denotes the cosine similarity of the negative pair within the same view, and the other is denotes the cosine similarity of the negative pair between different views;

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

[0037] (7),

[0038] wherein, denotes the calculation formula of the class cluster distribution probability;

[0039] The specific calculation formula of is as follows: ​​​

[0040] (8),

[0041] wherein, denotes the k-th cluster distribution of the v-th view, j'-th sample, denotes all cluster distributions of the v-th view.

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

[0043] (9),

[0044] wherein, and denote the original data of the view Figure 1 and the view Figure 2 .

[0045] Further, in the training process of jointly optimizing the multi-view reconstruction loss and the cross-view contrastive learning loss as the total task loss for model supervision, on the basis of relatively independent view reconstruction, the consistency between cross-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] wherein, denotes the multi-view reconstruction loss, denotes the cross-view contrastive learning loss.

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

[0049] In view of the problems that the existing work cannot fully utilize and fuse the multi-view information in the pre-learning process, and that the cross-view complementary enhancement relationship between the multi-view information is ignored in the representation pre-learning stage, which affects the clustering performance, the present application proposes a plug-in cross-view pre-learning method for multi-view clustering, which fully considers the cross-view complementary enhancement relationship between the multi-view information in the representation pre-learning stage, so as to improve the representation learning ability of the model and further optimize the clustering performance. Due to the excellent performance of contrastive learning in the consistency learning of view representation, the present application introduces the consistency information between the multi-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 representation learning, a cross-view supervision module is designed, which realizes the fusion of view reconstruction self-loss and cross-view contrast loss, considers the cross-view information while supervising the view self-feature learning, provides clearer task guidance for the model, and makes the guiding task of feature learning change from single reconstruction to more comprehensive consideration of multi-view semantic information, so that the learning of multi-view information is more accurate and targeted. Meanwhile, the plug-in cross-view pre-learning method for multi-view clustering provided by the application has high scalability and can be used with other clustering algorithms, providing flexible and effective solutions for different application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 The plug-in cross-view pre-learning method for multi-view clustering provided by the application is a flow chart.

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

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

[0054] The plug-in cross-view pre-learning method for multi-view clustering provided by the application, through the plug-in pre-learning method, before the clustering task is executed, that is, in the pre-learning part of the data, through the view reconstruction self-supervision module and the cross-view contrast supervision module, the representation learning of the multi-view information is introduced by using contrast learning, the features that can extract the information of the data itself and introduce the multi-view information are learned, so as to improve the performance of the downstream task, so that better clustering effect can be obtained under each clustering method.

[0055] The plug-in cross-view pre-learning method for multi-view clustering provided by the application mainly includes two parts: a multi-view feature learning module and a cross-view supervision module.

[0056] The multi-view feature learning module is mainly used to obtain the encoding representation of the original input data on each view, that is, the feature representation by using the multi-view autoencoder; the cross-view supervision module is mainly used to introduce the complementary consistency information between the multi-views by using the contrast learning method, and the multi-view reconstruction loss and the cross-view contrast learning loss are jointly optimized as the total task loss for model supervision.

[0057] 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 autoencoder, convolutional autoencoder, long short-term memory (LSTM) autoencoder, and adversarial autoencoder. In order to adapt to different data types, the present application uses a basic multi-view autoencoder to learn the representation of the original data.

[0059] In the multi-view clustering task, given a multi-view data set containing views and samples, and denoted as the original input data of view , where represents the input dimension of the original input data of view .

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

[0061] (1),

[0062] wherein, represents the encoder of view , represents the parameters corresponding to the encoder of view , represents the original input data of view , and represents the encoded representation of the original input data of view .

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

[0064] The cross-view supervision module aims to utilize the complementary consistency information between multi-views in a contrast learning manner to learn more discriminative view representations. The cross-view supervision module includes a view reconstruction self-supervision module and a cross-view contrast supervision module. The view reconstruction self-supervision module utilizes the intra-view reconstruction of each view to preserve the feature learning within the view; the cross-view contrast supervision module utilizes the complementary consistency information between multi-views to construct a cross-view contrast learning loss to promote multi-view representation learning.

[0065] Wherein, 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 encoded representation , the mathematical expression of its specific decoding process is:

[0068] (2),

[0069] wherein, represents a decoder, represents the parameters corresponding to the decoder, represents the reconstructed data of view . By narrowing the gap between the reconstructed data of view and the original input data of view , the neural network layer parameters that best express the data characteristics are learned to obtain the optimal encoded representation of view . Therefore, for a given view , its view reconstruction self-supervised loss function is defined as formula (3) :

[0070] (3),

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

[0072] Therefore, for a given multi-view data set containing views, the multi-view reconstruction loss is the sum of the reconstruction self-supervised losses of all views, that is:

[0073] (4),

[0074] The specific operation process of the cross-view contrast supervision module is as follows:

[0075] The cross-view contrast supervision module is used to introduce complementary consistency information between multi-views. For traditional multi-view data, a graph model can be used, wherein nodes represent text samples, and edges represent the relationship between samples. Multi-modal fusion can also be performed, including serial, parallel, attention mechanism, etc. By using the multi-modal fusion mechanism, the information between different views is integrated together. However, the above methods have many premise requirements and are not suitable for grafting in lightweight pre-trained models. Therefore, the present application uses a cross-view contrast learning method to introduce complementary consistency information between multi-views, that is, more semantic information is considered in the process of multi-view representation pre-training. Inspired by the idea that "labels are also a kind of representation", the present application regards the information facing the class cluster as different clustering distribution representation on each sample. Specifically, the present application uses two layers of nonlinear fully connected layers from the encoded representation data set of the original input data of view ​​​ The cluster distribution representation of the samples is extracted. For a given view The process can be formalized as:

[0076] (5),

[0077] wherein, denotes the number of class clusters in the data set, the i-th cluster distribution representation of the v-th view denotes the i-th cluster distribution representation in the v-th view denoted as . .

[0078] The present application uses the following loss function to learn the consistency of the maximum multi-view inter-class cluster distribution representation:

[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] wherein, denotes a temperature parameter for controlling the softness of the inter-class cluster distribution representation, which is fixed to 1.0 in the experiment. denotes the cosine similarity of the positive pair, which is formed by the cluster distribution representations of different views from the same class cluster. The calculation of the negative pair contains two cases, one is denotes the cosine similarity of the negative pair in the same view, and the other is denotes the cosine similarity of the negative pair between different views. In order to avoid the clustering results being mostly assigned to the same class cluster, the present application also considers the entropy of the class cluster assignment probability, and the corresponding calculation formula is as follows:

[0081] (7),

[0082] wherein, denotes the calculation formula of the cluster distribution probability;

[0083] The specific calculation formula of is as follows:

[0084] (8),

[0085] wherein, denotes the k-th cluster distribution of the v-th view and the j’-th sample, denotes all cluster distributions of the v-th view.

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

[0087] (9),

[0088] wherein, and represent the original data of view Figure 1 and view Figure 2 .

[0089] Then, joint optimization is performed to take 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] In the training process, compared with the traditional reconstruction loss, the application adopts the strategy of joint optimization, combines the multi-view reconstruction loss and the cross-view contrastive learning loss, and takes them as the task target of the model in the representation pre-learning stage. Specifically, on the basis of relatively independent view reconstruction, the consistency between cross 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, so that better clustering effect is achieved. The mathematical expression of the corresponding total task loss is:

[0091] (10),

[0092] wherein, represents the multi-view reconstruction loss, represents the cross-view contrastive learning loss. Through the joint optimization of the two, the supervision information in the view and between the views in the multi-view data can be considered comprehensively in the training process, the expression ability of the model in the multi-view data is effectively improved, the model is guided to learn the potential structure of the data, and thus the clustering performance is better improved.

[0093] As shown in Figure 1 , 2 indicates that the total loss is the weighted sum of the two losses, and the weight distribution can be specified artificially. The application takes the data set of two views as an example, and represent the original data of view Figure 1 and view Figure 2 , and represent the encoding representation, i.e. the feature representation, of the original data and , and represent the reconstruction data of view Figure 1 and view Figure 2 , so as to introduce in detail the plug-in cross-view pre-learning method for multi-view clustering. Specifically, for the given original data Figure 1 and Figure 2 of view and view , the corresponding encoding representation is obtained after encoding by the encoder and then the encoding representation and After decoding by the decoder, the reconstructed data of the view Figure 1 and the view Figure 2 and ; the multi-view reconstruction loss is calculated by using the view reconstruction self-supervision module; the cross-view contrast learning loss is calculated by using the cross-view contrast supervision module; then the multi-view reconstruction loss and the cross-view contrast learning loss are taken as the total task loss for model supervision.

[0094] The present application selects the MCJC benchmark model as the basis of the experiment, and designs the comparative experiment on each data set under the same environment and parameter configuration. Through careful comparative analysis, it is found that the pre-training representation has a great influence on the clustering performance, and sufficient pre-training representation makes the model show stable and efficient results on two key performance indicators of NMI and CA. In order to further verify the effect of the pre-training representation stage, the present application performs the pre-training representation stage with different iteration numbers on each data set, and the results are shown in Figure 2 and Figure 3 The experimental results on five different data sets (Movies, UCI, AMiner-700, AMiner, Reuters) are consistent, which shows that the model with sufficient pre-training has significant improvement in clustering accuracy and normalized mutual information. The performance improvement on the UCI and AMiner-700 data sets is particularly significant, which further confirms that sufficient and in-depth pre-training representation is crucial when dealing with data sets with high diversity and complexity.

[0095] ​Table 1 presents the clustering results comparison between the proposed multi-view clustering oriented plug-in cross-view pre-learning method and the prior art on various real data sets. PCRL_DEMVC, PCRL_MCJC, PCRL_CVCL and PCRL_ProImp show the effects of applying the proposed multi-view clustering oriented plug-in cross-view pre-learning method to different existing models. The results show that from the Movies, UCI, AMiner-700, AMiner to Reuters data sets, the proposed multi-view clustering oriented plug-in cross-view pre-learning method is superior to the MCJC benchmark model in two core indicators of 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, the proposed multi-view clustering oriented plug-in cross-view pre-learning method realizes cross-view contrast learning to integrate multi-view information, which can better capture the diversity of data, fully utilize the rich information in each view, provide clear task guidance for representation learning, and ensure the accuracy and relevance of the model when processing multi-view information.

[0096] Table 1 Clustering results comparison (%)

[0097]

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

[0099] The above only describes the preferred embodiments of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the principles of the present application, several improvements and refinements can be made, and these improvements and refinements should also be considered within the scope of protection of the present application.

Claims

1. A plug-in cross-view pre-learning method for multi-view clustering, characterized in that, The method comprises the following steps: First, a multi-view autoencoder is used to obtain the encoding representation of the original input data on each view; then, a cross-view contrastive learning method is used to introduce 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 to supervise the model; when the cross-view contrastive learning method is used to introduce complementary consistency information between multiple views, the class cluster-oriented information is regarded as different clustering distribution representations on each sample, and a two-layer nonlinear fully connected layer φ c The clustering distribution representation of the sample is extracted from the encoding representation dataset of the original input data from view v For a given view v, this process is formalized as: C v = φ c (Z v ), C v ∈ R N*K (5), where K denotes the number of class clusters in the dataset, C v the kth cluster distribution representation in view v, denoted as A loss function of formula (6) is used to learn the consistency of the maximum multi-view inter-class cluster distribution representation: where τ c denotes a temperature parameter, which is used to control the softness of the multi-view cluster distribution representation; denotes the cosine similarity of positive pairs, which is formed by cluster distribution representations of different views from the same cluster; the computation of negative pairs contains two cases, one is denotes the cosine similarity of negative pairs within the same view, and the other is denotes the cosine similarity of negative pairs between different views; The entropy of the class cluster assignment probability is considered to avoid the clustering results being mostly assigned to the same class cluster, and the corresponding calculation formula is as follows: wherein, represents a calculation formula of the cluster distribution probability; The specific calculation formula is as follows: wherein, denotes the k-th cluster distribution of the v-th view, the j'-th sample, c v denotes all cluster distributions of the v-th view; Therefore, the mathematical expression of the cross-view contrast learning loss is as follows: wherein X 1 and X 2 represent the original data of view 1 and view 2, respectively.

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 multi-view dataset containing V views and N samples, and denoted as denotes the original input data of view v, d v denotes the input dimension of the original input data of view v; for the original input data X v The mathematical expression of the learned representation corresponding to the encoding layer is: Z v = E v (X v ; φ v ) (1), where E v denotes the view v encoder, φ v denotes the parameters corresponding to the view v encoder, X v denotes the original input data of view v, Z v denotes the encoded representation of the original input data of view v.

4. The plug-in cross-view pre-learning method for multi-view clustering according to claim 3, characterized in that, For a given view v and its corresponding coded representation Z v The mathematical expression of its specific decoding process is: X v′ = D v Z v , φ v′ ) (2), where D v denotes the decoder, φ v′ denotes the parameters of the decoder, X v′ denotes the reconstructed data of view v; for a given view v, its view reconstruction self-supervised loss function is defined as in equation (3): wherein, represents the L2 norm formula, N represents the number of samples; For a given multi-view data set containing V views, the multi-view reconstruction loss is the sum of all view reconstruction self-supervised losses, that is:

5. The plug-in cross-view pre-learning method for multi-view clustering according to claim 1, characterized in that, When the cross-view contrast learning method is used to introduce complementary consistency information between multi-views, the intra-view reconstruction of each view is first used to retain the feature learning within the view, and then the complementary consistency information between multi-views is used to construct a cross-view contrast 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, In the training process of jointly optimizing the multi-view reconstruction loss and the cross-view contrast learning loss as the total task loss for model supervision, the consistency between cross-views is further considered on the basis of relatively independent view reconstruction, a corresponding cross-view contrast supervision target is designed, and the model parameters are optimized by minimizing the joint loss; the mathematical expression of the corresponding total task loss is as follows: L = L c + L r (10), where L r represents the multi-view reconstruction loss, L c represents the cross-view contrastive learning loss.

Images