Graph-based Multi-view Partial Multi-label Learning

Abstract

In multi-view partial multi-label learning (MVPML) problem, each instance is described by several heterogeneous feature representations and associated with a set of candidate labels, which include both ground-truth and noisy labels. The key to learn from MVPML data lies in how to deal with multi-view data and how to select the ground-truth labels from candidate label set. In this paper, we propose a Graph-based Multi-view Partial Multi-label method, which integrates exploiting multi-view information, noisy label disambiguation and training predictor model into a whole framework. Specifically, we first exploit the consensus information across different views by learning the similarity graph of each view and fuses these similarity graphs into a unified graph. Secondly, we decompose the observed label set into a ground-truth label matrix and a noisy label matrix, where the noisy label matrix is assumed to be sparse. Then, we embed the learned unified similarity graph into the process of label disambiguation to obtain a more reliable ground-truth label matrix. Finally, the predictive model is learned by the ground-truth label matrix. Extensive experiments indicate that our proposed method can achieve superior or comparable performance against state-of-the-art methods.