Advanced metrics for class-driven similarity search

Abstract

This paper presents two metrics for the nearest neighborclassifier that share the property of being adapted, i.e. learned, on aset of data. Both metrics can be used for similarity search when theretrieval critically depends on a symbolic target feature. The first oneis called local asymmetrically weighted similarity metric (LASM), and itexploits reinforcement learning techniques for the computation ofasymmetric weights. Experiments on benchmark datasets show that LASMmaintains good accuracy and achieves high compression ratesoutperforming competitor editing techniques like condensed nearestneighbor. The second metric, called the minimum risk metric (MRM), isbased on probability estimates. MRM can be implemented using differentprobability estimates and performs comparably to the Bayes classifierbased on the same estimates. Both LASM and MRM outperform the NNclassifier with the Euclidean metric