Classifier Calibration with ROC-Regularized Isotonic Regression

Code for experiments and figures of the paper "Classifier Calibration with ROC-Regularized Isotonic Regression", published at AISTATS 2024.

Files

• covtypesmall.csv: Dataset, exctracted from the UCI Covertype dataset.
• classifiers.py: Train logistic regression classifiers and save predictions on test and calibration set in the predictions/ folder.
• IRPexperiments.jl: Calibrate logistic regression with IRP and recursive binning for K=2,3,4. Log cross-entropy and AUC for calibration and test sets. Save figures in the figures/ folder.
• utils.jl: Julia functions to run our experiments.

Dependencies

• Python: Numpy, Pandas, Sklearn
• Julia: Random, Plots, PlotlyJS, Measures, LaTeXStrings, Polyhedra, GLPK, NPZ, Printf