Instance-Based Uncertainty Estimation for Gradient-Boosted Regression Trees

Authors: Jonathan Brophy, Daniel Lowd

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we find that IBUG achieves similar or better performance than the previous state-of-the-art across 22 benchmark regression datasets. We now compare IBUG s probabilistic and point predictions to NGBoost [22], PGBM [64], and CBU [44] on 21 benchmark regression datasets and one synthetic dataset.
Researcher Affiliation Academia Jonathan Brophy University of Oregon jbrophy@cs.uoregon.edu Daniel Lowd University of Oregon lowd@cs.uoregon.edu
Pseudocode Yes Algorithm 1 IBUG affinity computation. Algorithm 2 IBUG probabilistic prediction. Algorithm 3 IBUG accelerated tuning of k.
Open Source Code Yes Source code is available at https: //github.com/jjbrophy47/ibug.
Open Datasets Yes We now compare IBUG s probabilistic and point predictions to NGBoost [22], PGBM [64], and CBU [44] on 21 benchmark regression datasets and one synthetic dataset. Additional dataset details are in B.1. Links to all data sources as well as the code for IBUG and all experiments is available at https://github.com/jjbrophy47/ibug.
Dataset Splits Yes We use 10-fold cross-validation to create 10 90/10 train/test folds for each dataset. For each fold, the 90% training set is randomly split into an 80/20 train/validation set to tune any hyperparameters.
Hardware Specification Yes Experiments are run on publicly available datasets using an Intel(R) Xeon(R) CPU E5-2690 v4 @ 2.6GHz with 60GB of RAM @ 2.4GHz.
Software Dependencies No The paper mentions software like Python, Cython, XGBoost, Light GBM, and Cat Boost, but does not provide specific version numbers for these dependencies.
Experiment Setup Yes We tune NGBoost the same way as in Duan et al. [22]. Since PGBM, CBU, and IBUG optimize a point prediction metric, we tune their hyperparameters similarly. We also tune variance calibration parameters γ and δ for each method ( 3.2). Exact hyperparameter values evaluated and selected are in B.2.