Predictive Uncertainty in Gradient-Boosted Regression Trees : A Muon Energy Reconstruction Case Study
This repository hosts the code base for the report that I wrote during my internship at INPP, at NCSR Demokritos with Evangelia Drakopoulou and in the context of the ANNIE Experiment of Fermilab.
This project investigates the performance of various regression models (mainly Gradient Boosted Trees) in reconstructing muon energy from detector data. The focus is on evaluating and visualizing predictive uncertainty, using ensemble methods and statistical confidence intervals.
The full analysis, methodology and figures are detailed in the following report:
👉 Predictive_Uncertainty_in_Gradient_Boosted_Regression_Trees.pdf
To create a conda environment with the needed dependecies run:
conda env create -f env.yml
Then:
conda activate reco_env and
pip install ibug
To reproduce table 1 and train uncertainty models for tables 2 and 3 run bash reproduce_train.sh. The results for the table 1 will be in the 'results' directory.
To reproduce the metrics of each uncertainty model run one-by-one:
python ibug_catb_test.py
python ibug_xgb_test.py
python CBU_pred.py
python ibug_cbu.py