Reduced-Rank Local Distance Metric Learning

We propose a new method for local metric learning based on a conical combination of Mahalanobis metrics and pair-wise similarities between the data. Its formulation allows for controlling the rank of the metrics’ weight matrices. We also offer a convergent algorithm for training the associated model. Experimental results on a collection of classification problems imply that the new method may offer notable performance advantages over alternative metric learning approaches that have recently appeared in the literature.