Title: Null Hypothesis Test for Anomaly Detection
Abstract: In this talk we present a hypothesis test designed to exclude thebackground-only hypothesis for Anomaly detection searchs. Extending ClassificationWithout Labels, we show that by testing for statistical independence of the twodiscriminating dataset regions, we are able exclude the background-onlyhypothesis without relying on fixed anomaly score cuts or extrapolations ofbackground estimates between regions. The method relies on the assumption ofconditional independence of anomaly score features and dataset regions, whichcan be ensured using existing decorrelation techniques. As a benchmark example,we consider the LHC Olympics dataset where we show that mutual informationrepresents a suitable test for statistical independence and our method exhibitsexcellent and robust performance at different signal fractions even in presenceof realistic feature correlations.
For help, please contact Webmaster.