Incomplete Cross-modal Clustering Analysis

In the era of big data,data presents the heterogeneous characteristics,which could provide richer and more comprehensive information for data analysis.As a basic analysis tool for multi-source data,cross-modal clustering has become a research hotspot because it does not require labeling information.Cross-modal clustering aims to cluster the high-similar crossmodal data into the same group through the exploration and analysis of the relationship between different modalities.Most of existing cross-modal clustering algorithms all handle complete data,however,in practical applications,it could not ensure that all modalities are complete due to potential factors such as human,environment,equipment and so on.At present,existing incomplete data clustering methods are to learn complete subspaces or complete incomplete data,but they ignore the discriminant information of the learned clustering representation,which hinders the further improvement of clustering performance;What's more,these algorithms are unidirectional clustering models,thus ignore the reusability of the clustering results and the category information hidden among the data itself.Therefore,to handle the above problems,this paper proposes the incomplete cross-modal data clustering algorithm.The specific content is as follows:(1)Aiming at the problem that existing methods ignore the discrimination of the learning complete subspace.This paper proposes a shared-discriminative incomplete cross-modal data clustering framework(iCmSC:Incomplete Cross-Modal Subspace Clustering,iCmSC),which uses the idea of deep canonical correlation analysis to explore and mine the correlation between incomplete cross-modal data to obtain a consistent subspace clustering representation;At the same time,iCmSC adds l12-norm regularization to the consistent clustering representation to enhance the discriminative performance of the clustering representation and further improve the accuracy of incomplete cross-modal data clustering.(2)Aiming at the problem that the existing algorithms ignore the hidden category information among data itself and cause the outflow of important information,a self-supervised learning strategy for incomplete data(Partial Multi-Modal Clustering via Self-Supervised Multi-Modal CCA Network,PMC-SCCA)is proposed in this paper.The proposed method establishes a self-supervised cyclic feedback learning network,which feedbacks the clustering results to the subspace feature learning process and self-expression learning process to effectively mine the category information hidden in the data itself.Based on the previous research content,this paper expands incomplete cross-modal data to incomplete multi-modal data clustering and makes full use of the more comprehensive and richer information of multi-modal data,which greatly improves the performance of incomplete multi-modal data clustering.