Co-clustering based classification of multi-view data

Multi-view learning is an attractive area of research where data is represented using multiple views, each containing some useful information. Many multi-view learning algorithms use a unified objective function that needs to be optimized simultaneously to achieve a common goal, using a similarity or distance measure, such as the Euclidean or Cosine metric. Having a unified objective function, however, does not allow independent learning from the different views which can later be integrated to improve the overall learning experience. In this paper, we explore the principles of multi-view learning, i.e., complementary and consensus principle, to propose a new multi-view classification algorithm that simultaneously allows for both independent and unified learning. The complementary principle says that each source of data contains some information that is missing in other views while the consensus principle aims to maximize the agreement across the multiple views. Our proposed technique treats the data both individually as well as a part of a multi-view data. Within a view, we exploit a supervised co-clustering-based measure that independently learns similarity between instances of that view. Across views, these similarlity values are then shared using transfer-learning. The learning process involves both co-clustering and multi-view learning within a supervised learning framework. The test data is predicted independently on each view and a concensus based prediction across the views is used for the final label. Individually, each of supervised learning, co-clustering, and multi-view learning has been used in the literature, each with its own set of advantages. To the best of our knowledge, the current study is the premier attempt to combine these concepts for the classification task. Results show that the proposed approach significantly outperforms other recent and state-of-the-art algorithms on several sparse and high-dimensional datasets using the accuracy and normalized mutual information criteria.