Deep multi-view semi-supervised clustering with sample pairwise constraints

Multi-view clustering has attracted much attention thanks to the capacity of multi-source information integration. Although numerous advanced methods have been proposed in past decades, most them generally overlook significance weakly-supervised and fail preserve feature properties multiple views, thus resulting unsatisfactory performance. To address these issues, this paper, we propose a novel Deep Semi-supervised Clustering (DMSC) method, which jointly optimizes three kinds losses during networks finetuning, including multi-view loss, semi-supervised pairwise constraint loss autoencoders reconstruction loss. Specifically, KL divergence based is imposed on common representation data perform heterogeneous optimization, weighting prediction simultaneously. Then, innovatively integrate constraints into process by enforcing learned must-link samples (cannot-link samples) be similar (dissimilar), such that formed architecture can more credible. Moreover, unlike existing rivals only encoders for each branch further tune intact frame contains both decoders. In way, issue serious corruption view-specific view-shared space could alleviated, making whole training procedure stable. Through comprehensive experiments eight popular image datasets, demonstrate our approach performs better than state-of-the-art single-view competitors.