Dual distance adaptive multiview clustering

Abstract

Adaptively adjusting the graph, by taking the clustering capability into consideration, has become popular in graph-based clustering methods and extended to multiview clustering problem. Existing methods learn the graph from pairwise distances of the data in the original space or a linearly projected space, which requires that those representations can finely reflect the implicit data structure. However, the data structure in high-dimensional space may not always lie on a linear manifold, and this problem becomes more critical in multiview conditions. Aim at this, we propose a multiview clustering method based on adaptive graph and dual distance. Specifically, we fuse the distances computed from nonlinear embedding space and original space. An adaptive graph is then constructed based on the fused distance, and it is more reliable for multiview clustering. In our approach, the clustering result with exact number of clusters can be found without post-processing. We evaluate the proposed method on several multiview datasets, the experimental results show our approach is superior to the state-of-the-art multiview clustering methods.