Layered, Multivariate Anomaly Explanations: A First Look [pdf]
At the International Workshop on Statistical Relational Learning, (SRL-2009).

Matthew Michelson and Sofus A. Macskassy

Abstract

An anomaly is a data point that deviates dramatically from some set of related data points, based on some common metric. For instance, consider a set of flights. Some might be anomalous because their flying time is so much larger than the average even though they fly to the same sets of airports. While there is signicant research on finding anomalous records from within data sources, we are unaware of research on explaining why those anomalous data points are actually anomalous (beyond their deviance from the "standard" value for the metric). Generating these explanations is the focus of this paper.