Exploring confidence-based neighborhoods in Outlier Detection

This paper presents a non-traditional approach to detect outliers based on local information captured from neighbors. The Confidence-based Outlier Detection (COD) approach is proposed to explore the neighborhood for each target sample in order to obtain high detection confidence performed without being affected by irrelevant ambiguous data. In other words, the adopted SVM classifier is generated using k nearest neighbors and the value of k is decided based on the LDOF local density and SVM detection confidence. It is noted, the number of nearest neighbors could be incrementally explored, and it is not necessary to define the neighborhood size in advance. Our experimental results are presented that the proposed COD is effective to explore the specific neighborhood for each target sample and is able to obtain higher accuracy in outlier detection than other baseline approaches.