Passive-aggressive online distance metric learning and extensions

Abstract

Online metric learning aims at constructing an appropriate dissimilarity measure from training data composed of labeled pairs of samples. Margin maximization has been widely used as an efficient approach to address this problem. This paper reviews various existing online metric learning formulations, and also introduces an alternative passive-aggressive scheme. In addition, the pros and cons of each alternative are analyzed in a comparative empirical study over several databases. © 2013 Springer-Verlag Berlin Heidelberg.