Revisiting attribute independence assumption in probabilistic unsupervised anomaly detection

In this paper, we revisit the simple probabilistic approach of unsupervised anomaly detection by estimating multivariate probability as a product of univariate probabilities, assuming attributes are generated independently. We show that this simple traditional approach performs competitively to or better than five state-of-the-art unsupervised anomaly detection methods across a wide range of data sets from categorical, numeric or mixed domains. It is arguably the fastest anomaly detector. It is one order of magnitude faster than the fastest state-of-the-art method in high dimensional data sets.