UCSC-SOE-23-12: Anomaly Detection in Extremes

Peter Trubey and Bruno Sansó
10/24/2023 11:40 AM
We develop anomaly detection scores leveraging the independence between the radial and angular components of vectors in extreme value settings. The angular density is modeled as a Bayesian non-parametric mixture of projected gammas. The resulting posterior predictive density is used for the angular score. For flexible cate- gorical data modeling, we develop an extension of the projected gammas model using a Dirichlet-multinomial kernel. This is coupled with our proposed anomaly detec- tion score in mixed data settings. We evaluate the anomaly detection efficacy of our proposed scores and find that they are generally superior to tested canonical methods.