Informed multimodal latent subspace learning via supervised matrix factorization

Abstract

Matrix factorization technique has been widely used as a popular method to learn a joint latent-compact subspace, when multiple views or modals of objects (belonging to single-domain or multiple-domain) are available. Our work confronts the problem of learning an informative latent subspace by imparting supervision to matrix factorization for fusing multiple modals of objects, where we devise simpler supervised additive updates instead of multiplicative updates, thus scalable to large scale datasets. To increase the classification accuracy we integrate the label information of images with the process of learning a semantically enhanced subspace. We perform extensive experiments on two publicly available standard image datasets of NUS WIDE and compare the results with state-of-the-art subspace learning and fusion techniques to evaluate the efficacy of our framework. Improvement obtained in the classification accuracy confirms the effectiveness of our approach. In essence, we propose a novel method for supervised data fusion thus leading to supervised subspace learning.