Localizing SVM using an adaptive neighborhood distance

One of the most important problems of the nearest neighbor and related classifiers is the distance measure. The distance measure is the fundamental part to compute the neighbors of a test instance. Using the nearest neighbors as the training instances of another classifier is a usual form of localizing a classifier such as SVM. In this paper, a method is proposed to adapt the distance measure by weighting instances in order to improve the performance of Local-SVM classifier. A positive weight is assigned to each instance and instances that have no influence on the performance of SVM get the weight 0, and therefore will be removed in the training phase. The proposed method found a local optimal solution and weights the instances in order to maximize Leave-One-Out performance of Local-SVM.