Bayesian Anomaly Detection (BAD v0.1)

Prior experiments with Bayesian rule generation produced a scalable anytime learner. At its core, that tool computes the likelihood of new events as the product of frequencies of old events. Orrego and Menzies applied that tool to logs of an F-15 flight simulator and showed that the same tool can detect anomalous events which have not been seen previously. This paper checks the external validity of that prior experiment. In twenty-five data sets, anomalous new situations could be identified with high probabilities of detection (average pd over 80%) and low probabilities of false alarm (usually, pf ≤ 5%). These results strongly suggest that we can detect anomalous events, even among very large data sets.