Adaptive Multiple-View Label Propagation for Semi-supervised Classification

Abstract

We propose a novel Adaptive Multiple-view Label Propagation (MLP) framework for semi-supervised classification. MLP performs classification on multiple views rather than on the single view, and can exploit the complementarity of multiple views in the label prediction process. Moreover, MLP integrates the multi-view label propagation and the adaptive multiple graph weight learning into a unified model, where a linear transformation is used to enforce different weights form different view spaces. Thus, an optimal graph weight matrix can be constructed from each view. Due to the adaptive manner for defining the weight matrices from multiple views, MLP can avoid the complex and tricky process to select the neighborhood size or kernel parameter. Extensive results on real data show that MLP can deliver the enhanced performances, by comparing with other related techniques.