Risk-based adaptive metric learning for nearest neighbour classification

Abstract

The performance of k-nearest neighbour classification highly depends on the appropriateness of distance metric designation. Optimal performance can be obtained when the distance metric is matched to the characteristics of data. Existing works on distance-metric learning typically learn a global linear transform from training samples, and the effectiveness is limited to data, which are well-separated by linear decision boundaries. To address this problem, we propose a locally adaptive weighted distance-metric learning method to deal with the non-linearity of the data. The metric are learned based on local leave-one-out cross-validation (LOOCV) risks in each dimension, so that the local variations in feature component discriminability are taken into account. Experiments on both public datasets and hyper-spectral imagery classification demonstrate that the classification accuracy of the proposed method shows about 2–10% improvements over other competitive methods.