Similarity Kernels for Nearest Neighbor-Based Outlier Detection

Outlier detection is an important research topic that focuses on detecting abnormal information in data sets and processes. This paper addresses the problem of determining which class of kernels should be used in a geometric framework for nearest neighbor-based outlier detection. It introduces the class of similarity kernels and employs it within that framework. We also propose the use of isotropic stationary kernels for the case of normed input spaces. Two definitions of similarity scores using kernels are given: the k-NN kernel similarity score (kNNSS) and the summation kernel similarity score (SKSS). The paper concludes with preliminary experimental results comparing the performance of kNNSS and SKSS for outlier detection on four data sets. SKSS compared favorably to kNNSS.