Smart Sampling: A Novel Unsupervised Boosting Approach for Outlier Detection

Abstract

While various ensemble algorithms have been proposed for supervised ensembles or clustering ensembles, there are few ensemble based approaches for outlier detection. The main challenge in this context is the lack of knowledge about the accuracy of the outlier detectors. Hence, none of the proposed approaches focused on sequential boosting techniques. In this paper for the first time we propose a novel boosting algorithm for outlier detection called BSS, where we sequentially improve the accuracy of each ensemble detector in an unsupervised manner. We discuss the effectiveness of our approach in terms of bias-variance trade-off. Furthermore, an extended version of BSS (called DBSS) is proposed to introduce a novel source of diversity in outlier ensemble modeling. DBSS is used to analyze the effect of changing the input parameter of BSS on its detection accuracy. Our experimental results on both synthetic and real data sets demonstrate that our approaches outperform the two state-of-the-art outlier ensemble algorithms and benefit from bias reduction. In addition, our BSS approach is robust with respect to the changing input parameter. Since each detector in our proposed BSS/DBSS is only a subset of the whole dataset, our both techniques are well suited to application environments with limited memory processors (e.g., wireless sensor networks).