Smooth representation learning from multi-view data.

Multi-view subspace clustering has aroused more and more attention due to its ability to explore data correlation from multiple views without stressful label annotations. Although a plethora of methods have been developed, they are powerless considering the original data may not be separable into subspaces. In this paper, we target to achieve a smooth representation from multi-view data which is committed to facilitating the downstream clustering task. Our assumption is that samples in the same cluster always tend to be densely connected. In detail, the proposed method can maintain the graph geometric structure by performing graph filtering to obtain the smooth representation. Furthermore, we unify the smooth representation learning and the subsequent multi-view clustering in a joint framework, hence the "multi-view clustering-friendly"representation can be expected. As a result, the smooth representation learning of each view and the achieving of multi-view clustering can be boosted in a mutual reinforcement manner towards an overall optimal solution. Extensive experiments on common multi-view datasets are conducted to demonstrate the effectiveness of our mode, compared to other SOTAs over the clustering performance.