Mining best strategy for multi-view classification

Abstract

In multi-view classification, the goal is to find a strategy for choosing the most consistent views for a given task. A strategy is a probability distribution over views. A strategy can be considered as advice given to an algorithm. There can be several strategies, each allocating a different probability mass to a view at different times. In this paper, we propose an algorithm for mining these strategies in such a way that its trust in a view for classification comes close to that of the best strategy. As a result, the most consistent views contribute to multi-view classification. Finally, we provide experimental results to demonstrate the effectiveness of the proposed algorithm.