Semi-supervised classification with metric learning

Abstract

Metric learning performs a task of constructing a metric space that reflects relationship of training data. Both supervised and semi-supervised settings are well studied. In this paper, we propose a method to perform semi-supervised classification in a metric learning setting. The proposed method is based on non-metric Multi-Dimensional Scaling (NMDS). An original metric space is generated using labeled data by NMDS. Unlabeled data is added to this metric space and an updated procedure is used to maintain the consistence of the space. This method deals with unlabeled points one by one compared to the traditional label propagation method in semi-supervised learning setting. Also in the proposed method, we use property of local consistence of Euclidean Distance to get a fair reasonable result. Our method avoids pure Euclidean Distance description of original data representation. The proposed method is applied to UCI beach mark data sets and experimental results show that it is effective. © 2010 IEEE.