This repository explores the possibility of applying the DCTR technique to anomaly detection. The setting is that there is some feature m where the signal is expected to be resonant. There are other features x that can be used to enhance signal over background. In the CWoLa hunting methodology, one can only use features x that don't sculpt a bump (a looser requirement than being independent from m, but clearly that would be sufficient); then a parameteric fit based on the sidebands is used to estimate the background in the signal region. In this new approach, a DCTR model is used to reweight simulation to data in away from the signal region. This model is parameterized in m and then interpolated to the signal region. The reweighted simulation in the signal region can then be used to make a classifier (using x) as well as estimate the background. There is no need for a sideband fit if the reweighting works well.
We provide a few example notebook to illustrate how DCTR Hunting works
To illustrate the basic idea behind DCTR hunting, we provide
- A basic example just using a Gaussian distribution
./Toy.ipynb
To be filled in
- Based on the LHCO2020 dataset.