Exact Learning and Data Compression with a Local Asymmetrically Weighted Metric

This paper is concerned with a local asym-metric weighting scheme for the nearest neighbor classiication algorithm and a learning procedure, based on reinforcement, for computing the weights. Theoretical results show that this context dependent metric can learn exactly certain classes of concepts storing fewer examples that those required by the Euclidean metric. Moreover, computer experiments show that the proposed metric, storing a randomly chosen small percentage of the training set, can still outperform the standard nearest neighbor algorithm in terms of accuracy and can obtain almost the same accuracy as the k-NN algorithm. The drastic reduction of the stored examples results in a great speed up of performance at query time.