Clustering Based on MultiView Diffusion Maps

Ofir Lindenbaum, Arie Yeredor, Amir Averbuch

Abstract: We consider a reduced dimensionality representation based on multiple views of the same underlying process. These multiple views can be obtained, for example, using several different modalities, measured with different instrumentation or generated based on different methods of feature extractions. Our framework is based on a cross-view random walk process which is restrained to hop between the different views in each time step. The random walk model is constructed using the intrinsic relation within each view as well as the mutual relations between views. Within this framework, multiview diffusion distances are defined which lead to reduced representations for each view. The reduced representations are exploited to perform clustering. The applicability of the multiview approach for clustering is demonstrated on both artificial and real data.