Less is More: Building Selective Anomaly Ensembles

摘要: Ensemble learning for anomaly detection has been barely studied, due to difficulty in acquiring ground truth and the lack of inherent objective functions. In contrast, ensemble approaches classification clustering have studied effectively used long. Our work taps into this gap builds a new approach detection, with application event temporal graphs as well outlier no-graph settings. It handles combines multiple heterogeneous detectors yield improved robust performance. Importantly, trusting results from all constituent may deteriorate overall performance ensemble, some could provide inaccurate depending on type data hand underlying assumptions detector. This suggests that combining selectively is key building effective ensembles—hence “less more”.In paper we propose novel called SELECT which automatically systematically selects combine fully unsupervised fashion. We apply our method multi-dimensional point (no-graph), where successfully utilizes five base seven consensus methods under unified framework. extensive quantitative evaluation real-world datasets (four events), including Enron email communications, RealityMining SMS phone call records, New York Times news corpus, World Cup 2014 Twitter feed. also UCI Machine Learning Repository. Thanks its selection mechanism, yields superior compared individual alone, full (naively results), an existing diversity-based weighted approach.