Unified Bayesian Anomaly Detection and Classification

14.05.2018 11:00 – 12:00

Although there are many machine learning algorithms for classification and anomaly detection, these are not statistically rigorous and are typically handled by different algorithms. As a result they are sub-optimal for the needs of fundamental science experiments such as the LSST and SKA. Here we present a unified and fully Bayesian Anomaly Detection and Classification framework that satisfies these needs. The method is novel in that it accounts for measurement error and statistical variation on observations. Meaningful classification probabilities are important, since what we are actually after is inference of physical parameters. By marginalising over the object type/classification we achieve this without needing to decide on a threshold for producing hard labels. We demonstrate this method on multiple synthetic datasets and compare results with those from Random Forest, IsolationForest and Local Outlier Factor (LOF), showing that we achieve superior performance in nearly all cases.

Lieu

Bâtiment: Ecole de Physique

Salle 234, 24 quai Ernest-Ansermet

entrée libre