FETA: Flow-Enhanced Transportation for Anomaly Detection

This is the codebase for the FETA (Flow-Enhanced Transportation for Anomaly Detection) method for background construction.

Accompanying paper: https://arxiv.org/abs/2212.11285 (authors: Tobias Golling, Samuel Klein, Radha Mastandrea, Benjamin Nachman)

For questions/comments about the code contact: [email protected]

Pipeline

Prepare and preprocess the data.

Use the notebook generate_dataset_and_preprocess.ipynb.

Generate the flow and train it to map between simulation and data in the SB.

Use the script run_full_cycle_2step.py. The script also applies the trained flow to data in the SR to generate the background template samples.

Compare the FETA-generated background template samples with SR data.

Use the script final_eval_SR.py.

Generate all results plots.

Use the notebook make_sig_rej_plots.ipynb.

Other comments

The folder auxiliary plots contains code to generate the "prettier" schematics.

To make the scatterplots of classifier scores for the different background construction methods (Figs 10 and 11 in the main paper), run run_scatter_plot.py to process the data, then analyze_scatterplot.ipynb to make the visualizations.

Optimal Transport studies

Accompanying report: https://arxiv.org/abs/2212.06155 (authors: Radha Mastandrea, Benjamin Nachman)

The folder full_cycle_scripts contains alternate versions of the run_full_cycle_2step.py script that were used in the supplementary optimal transport studies. The notebook OT_studies.ipynb contains the code to make all of the visualizations (from the data generated from run_full_cycle_xx.py) shown in that paper.