Output Regularized Metric Learning with Side Information

Distance metric learning has been widely investigated in machine learning and information retrieval. In this paper, we study a particular content-based image retrieval application of learning distance metrics from historical relevance feedback log data, which leads to a novel scenario called collaborative image retrieval. The log data provide the side information expressed as relevance judgements between image pairs. Exploiting the side information as well as inherent neighborhood structures among examples, we design a convex regularizer upon which a novel distance metric learning approach, named output regularized metric learning, is presented to tackle collaborative image retrieval. Different from previous distance metric methods, the proposed technique integrates synergistic information from both log data and unlabeled data through a regularization framework and pilots the desired metric toward the ideal output that satisfies pairwise constraints revealed by side information. The experiments on image retrieval tasks have been performed to validate the feasibility of the proposed distance metric technique.