Dual Graph-Regularized Multi-view Feature Learning

Abstract

Real-world datasets often describe data instances in different views that complement information for each other. Unfortunately, synthesizing these views for learning a comprehensive description of data items is challenging. To tackle it, many approaches have been studied to explore correlations between various features by assuming that all views can be projected into a same semantic subspace. Following this idea, we propose a novel semi-supervised method, namely dual graphregularized multi-view feature learning (DGMFL), for data representation in this paper. The core idea is to generate a latent subspace among different views. Our approach utilizes dual graph regularization to capture semantic relationships among data items on both multi-view features and label information, as well as locates view-specific features for each view to reduce the effects of uncorrelated items. In this way, DGMFL could achieve more comprehensive representations hidden in multi-view datasets. Extensive experiments demonstrate that DGMFL model is superior to state-of-the-art multi-view learning methods on real-world datasets