Multi-view Semantic Learning for Data Representation

Abstract

Many real-world datasets are represented by multiple features or modalities which often provide compatible and complementary information to each other. In order to obtain a good data representation that synthesizes multiple features, researchers have proposed different multi-view subspace learning algorithms. Although label information has been exploited for guiding multi-view subspace learning, previous approaches either fail to directly capture the semantic relations between labeled items or unrealistically make Gaussian assumption about data distribution. In this paper, we propose a new multi-view nonnegative subspace learning algorithm called Multi-view Semantic Learning (MvSL). MvSL tries to capture the semantic structure of multi-view data by a novel graph embedding framework. The key idea is to let neighboring intra-class items be near each other while keep nearest inter-class items away from each other in the learned common subspace across multiple views. This nonparametric scheme can better model non-Gaussian data. To assess nearest neighbors in the multi-view context, we develop a multiple kernel learning method for obtaining an optimal kernel combination from multiple features. In addition, we encourage each latent dimension to be associated with a subset of views via sparseness constraints. In this way, MvSL is able to capture flexible conceptual patterns hidden in multi-view features. Experiments on two real-world datasets demonstrate the effectiveness of the proposed algorithm.