Enhanced local support vector machine with fast cross-validation capability

Local SVM is a lazy learner combining k-nearest neighbor search and support vector machine classifier. We propose an improved implementation of local SVM which utilizes tree structure for efficient nearest neighbor search and a method to avoid unnecessary SVM training in areas far from decision boundary. The proposed lazy learner has great advantage on cross-validation efficiency while maintaining comparable accuracy to traditional SVM. The proposed method also enables us to conduct leave-one-out cross-validation which is previously considered too time-consuming to be practical on large dataset.