Chunk incremental distance metric learning algorithm based on manifold regularization

Abstract

In many real-time applications, observed samples always arrive in the form of chunks stream, traditional batch distance metric algorithms can hardly work well in such scenarios. This paper proposes a novel semi-supervised chunk incremental metric learning algorithm on the basis of the pairwise constraints. One general model is given to learn metric incrementally on the arriving chunks at first with its limitation of over-fitting overcame by utilizing extended constraint sets. Then, a manifold regularization term is used to keep locality adjacency structure of chunks during metric learning. Experimental results indicate superiorities of our algorithm, which obtains better accuracy and lower computation costs than existing incremental metric learning algorithms, and needs much less storage costs than batch ones.