yandex-research/GBDT-uncertainty

Uncertainty in Gradient Boosting via Ensembles

This is a supplementary code for our paper: "Uncertainty in Gradient Boosting via Ensembles" by Andrey Malinin, Liudmila Prokhorenkova, Aleksei Ustimenko (ICLR 2021)

See also our tutorials on uncertainty estimation with CatBoost: blog post with synthetic regression example, blog post with practical classification example.

Datasets can be found here.

Training models

python train_models.py regression 1

First argument options: regression, classification, regression_rf, classification_rf
Second argument (for CatBoost only): 0 or 1 indicates whether to tune hyperparameters (or use already obtained ones)

Aggregating results and getting tables in latex format

Regression:

python aggregate_results_regression.py prr_auc

Options: std, nll_rmse, prr_auc, rf_rmse, rf_prr_auc

Classification:

python aggregate_results_classification.py prr_auc

Options: nll_error, prr_auc, rf_nll_error, rf_prr_auc

Synthetic experiments

synthetic_regression.ipynb

synthetic_classification.ipynb

Additional experiment on KDD-99 Intrusion Detection dataset

(not included in the paper)

gbdt_uncertainty/kdd/kdd.sh