RKOF: Robust Kernel-Based Local Outlier Detection

Abstract

Outlier detection is an important and attractive problem in knowledge discovery in large data sets. The majority of the recent work in outlier detection follow the framework of Local Outlier Factor (LOF), which is based on the density estimate theory. However, LOF has two disadvantages that restrict its performance in outlier detection. First, the local density estimate of LOF is not accurate enough to detect outliers in the complex and large databases. Second, the performance of LOF depends on the parameter k that determines the scale of the local neighborhood. Our approach adopts the variable kernel density estimate to address the first disadvantage and the weighted neighborhood density estimate to improve the robustness to the variations of the parameter k, while keeping the same framework with LOF. Besides, we propose a novel kernel function named the Volcano kernel, which is more suitable for outlier detection. Experiments on several synthetic and real data sets demonstrate that our approach not only substantially increases the detection performance, but also is relatively scalable in large data sets in comparison to the state-of-the-art outlier detection methods.

This work is supported in part by the NSFC (Grant No. 60825204, 60935002 and 60903147) and the US NSF (Grant No. IIS-0812114 and CCF-1017828).