Weighted sparse simplex representation: a unified framework for subspace clustering, constrained clustering, and active learning

Abstract Spectral-based subspace clustering methods have proved successful in many challenging applications such as gene sequencing, image recognition, and motion segmentation. In this work, we first propose a novel spectral-based algorithm that seeks to represent each point sparse convex combination of few nearby points. We then extend the constrained active learning framework. Our motivation for developing framework stems from fact typically either small amount labelled data are available advance; or it is possible label some points at cost. The latter scenario encountered process validating cluster assignment. Extensive experiments on simulated real datasets show proposed approach effective competitive with state-of-the-art methods.