Neighbor k-convex-hull ensemble method based on metric learning
Mou Lian-ming · Journal of Hefei University of Technology · 2013
The k-local convex distance nearest neighbor classifier(CKNN) corrects the decision boundary of kNN when the amount of the training data is small,thus improving the performance of kNN.The k sub-convex-hull classifier(kCH) weakens the sensitivity of CKNN to the number of classes and the ring structure of samples distribution,hence improves the classification performance.But this method is still sensitive to the distance metric.Moreover,different types of samples in k nearest neighbors of a test instance are often seriously imbalanced,which leads to the decline of classification performance.In this paper,a neighbor k-convex-hull classifier(NCH) is proposed to address these problems.The robustness of the neighbor k-convex-hull classifier is improved by the techniques of metric learning and ensemble learning.Experimental results show that the proposed neighbor k-convex-hull classifier ensemble method,which is based on metric learning,is significantly superior to some state-of-the-art nearest neighbor classifiers.