Learning large margin nearest neighbor classifiers via cutting plane algorithm

The performance of popular and classical k-nearest neighbor classifier depends on the distance metric. Large margin nearest neighbor classifier using gradient optimization method is prone to local minima. In this paper, we present a Mahalanobis metric learning method based on cutting plane algorithm which reduces largely constraints for solving the semidefinite programming problem. Experimental results on the ITC I data sets show that our method can achieve promising speedups compared with the gradient based method under the similar training, test error rates.