SHARD: A Framework for Sequential, Hierarchical Anomaly Ranking and Detection

摘要

This work explores unsupervised anomaly detection within sequential, hierarchical data. We present a flexible framework for detecting, ranking and analyzing anomalies. The framework 1) allows users to incorporate complex, multidimensional, hierarchical data into the anomaly detection process; 2) uses an ensemble method that can incorporate multiple unsupervised anomaly detection algorithms and configurations; 3) identifies anomalies from combinations of categorical, numeric and temporal data at different conceptual resolutions of hierarchical data; 4) supports a set of anomaly ranking schemes; and 5) uses an interactive tree hierarchy visualization to highlight anomalous regions and relationships. Using both synthetic and real world data, we show that standard anomaly detection algorithms, when plugged into our framework, maintain a high anomaly detection accuracy and identify both micro-level, detailed anomalies and macro-level global anomalies in the data.