Local voting based multi-view embedding

This paper studies subspace based multi-view learning, investigating how to mine useful information within various kinds of data or features (views) and achieve the optimal cooperation between views. Unlike most existing methods focused on learning an optimal weighting scheme to linearly combine different types of view information, we propose to first improve the original information provided by each view by designing a voting based scheme to model individual neighbor structures of the data. This leads to a set of refined local proximity matrices corresponding to different confidence levels. Then, different schemes can be applied to further combine this refined set of composite local neighborhood representations. Also, we provide the semi-supervised version of the proposed algorithms to incorporate partially labeled objects. The experimental results demonstrate effectiveness and robustness of the proposed algorithms.