Efficient and interpretable ensemble methods for outlier detection

Abstract

Outlier detection aims at finding instances in a data set that do not conform to normal behavior. Ensemble methods have been recently studied for unsupervised outlier detection in highdimensional data sets. The critical assumption of ensemble methods is that the combination of several weak detectors (base models) can form a robust detector with better performance. In this thesis, we analyze the strengths and characteristics of some recent ensemble outlier detectors. We utilize the theory of locality-sensitive hashing to propose a generic ensemble framework named LSH iTables which can instantiate the most recent ensemble methods. We show that our framework has a strong connection to standard distance-based outlier detection compared to the other ensemble methods. Empirically, LSH iTables with suitable parameters can achieve a competitive performance on detection accuracy and method efficiency on both static and dynamic data. It can also provide sensitivity histogram to enable outlier interpretations.