A Novel SVM-Based Reduced NN Classification Method

In pattern classification or machine learning, instance-based learning (IBL) has gained much attention and can yield superior performance in many domains. In IBL, however, the storage requirement is proportional to the number of training instances. Furthermore, it usually takes too much time to classify a new, unseen instance because all training instances need to be considered in determining the ´nearness´ or ´similarity´ among instances. This paper proposes a novel SVM-based instance selection method for reduced instance based learning. Support Vector Machine (SVM) is employed to form an instance space regarding all training instances of a particular pattern classification problem. Wrapper-based classification performance validation technique is applied to find the best hyper parameter for support vector identification. With high classification effectiveness, informative support vectors (instances) lying on the margin hyper plane in the instance space will be identified and selected. Thereafter, a reduced training set is derived for reduced nearest neighbor classification. In this manner, new, unseen instances can be classified by using a few training instances. Some datasets are used to demonstrate the performance of the proposed instance selection method. Experimental results indicate that the classification accuracy can be maintained or even increased with a small number of remained training instances.