Anomaly Detection Ensembles: In Defense of the Average

When given multiple models it is often useful to combine them for improved reliability or performance over the individual models. Over the years many outlier metrics and detection methods have been developed for the purposed of finding data incongruous with the rest of the data. Inspired by the successes of supervised ensemble machine learning, many have proposed combining multiple anomaly detection methods together. We investigate the usefulness of building ensembles for the purpose of anomaly detection. We find that currently, to the best of our knowledge, there is no great advantage in using anything more complicated than the simple average over all available outlier scores.