Incomplete Multi-view Contrastive Learning

In the era of big data,multi-view data widely exists in the real world.For exam-ple,on the Internet,the data of web pages can be described by the feature set of the information contained in the web page itself,or by the information contained in the hy-perlink.Different descriptions constitutes multiple views of things.Compared with the traditional single-view data,multi-view data contains more information and can better describe the relationship between data,so the research on multi-view data has impor-tant value and practical significance,as a result,multi-view learning and unsupervised multi-view clustering are developed.The important premise of multi-view clustering and multi-view learning is that the view,the data in the view and the corresponding relationship between data points are complete.However,in practice,due to the com-plexity of data collection and transmission,data loss is common.This leads to the so-called incomplete multi-view problem.In order to solve the problem of incomplete multi-view,scholars at home and abroad have proposed some incomplete multi-view clustering algorithms.The goal of these algorithms is how to combine and utilize the feature information between different view data to reduce the impact of missing view sample data and improve clustering performance to approach the clustering effect in the complete data scenarios.For these works,this paper summarizes two challenging questions:(1)How to learn informative and consistent representations across different views without the help of labels.(2)How to restore the lost views the from existing data.In order to solve these challenging problems,in this paper,we combine cross-view consistent learning and missing view recovery into a unified framework from the per-spective of information theory.Based on this framework,we propose a noval method-Incomplete Multi-view Contrastive Learning(IMCL),which is applied to the incom-plete multi-view clustering problem,and data augmentation and contrastive learning are innovatively applied to the proposed method.Data augmentation makes the in-put of the model more diversified and has stronger generalization ability.Contrastive learning enhances the cross-view consistent learning,and there is no need to use neg-ative pairs for comparison,which avoids unstable training and improves the training efficiency of the model.The IMCL method firstly simulates the scene where the view data is missing by data augmentation,and then expands the view data by data augmenta-tion to add new training data sets.Next,the augmented view data is input into the same model network and consistency learning is carried out with the original missing view data to obtain the potential representation with more information.At the same time,we design and improve the joint optimization model network of four modules objective functions,which are within-view consistency learning loss,within-view reconstruction loss,cross-view contrastive learning loss and cross-view dual prediction loss.In brief,the IMCL method learns informative and consistent representations by maximizing the mutual information between different views through contrastive learning,and then re-covers the missing view data by minimizing the conditional entropy of different views through dual prediction.The proposed method is compared with eleven comparative multi-view cluster-ing methods on four widely used multi-view data sets.Extensive experimental results show the clustering performance of the proposed method remarkably outperforms the contrastive multi-view clustering methods,and prove the effectiveness of the proposed method.