Hierarchical Shrinkage: Improving the accuracy and interpretability of tree-based models.

Authors: Abhineet Agarwal, Yan Shuo Tan, Omer Ronen, Chandan Singh, Bin Yu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments over a wide variety of realworld datasets show that HS substantially increases the predictive performance of decision trees, even when used in conjunction with other regularization techniques.
Researcher Affiliation Academia 1Department of Statistics, UC Berkeley, Berkeley, California, USA 2EECS Department, UC Berkeley, Berkeley, California, USA. Correspondence to: Bin Yu <binyu@berkeley.edu>.
Pseudocode No The algorithm is described in detail using mathematical formulas (Equation 1 and 2) and text within Section 2, but there is no distinct block labeled 'Pseudocode' or 'Algorithm'.
Open Source Code Yes All code and models are released in a fullfledged package available on Github. HS is integrated into the imodels package github.com/csinva/imodels (Singh et al., 2021) with an sklearncompatible API.
Open Datasets Yes In this section, we study the performance of HS on a collection of classification and regression datasets selected as follows. For classification, we consider a number of datasets used in the classic Random Forest paper (Breiman, 2001; Asuncion & Newman, 2007), one (Breast cancer with id=13) from the open ML repository, as well as two (Juvenile and Recidivism) that are commonly used to evaluate rule-based models (Wang, 2019). For regression, we consider all datasets used by Breiman (2001) with at least 200 samples, as well as a variety of data-sets from the PMLB benchmark (Romano et al., 2020) ranging from small to large sample sizes.
Dataset Splits Yes In all cases, 2/3 of the data is used for training (hyperparameters are selected via 3-fold CV on this set) and 1/3 of the data is used for testing.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments, such as GPU or CPU models.
Software Dependencies No The paper mentions software packages like 'ranger', 'scikit-learn', 'bartpy', and 'imodels' but does not specify their version numbers, which are necessary for reproducible dependency descriptions.
Experiment Setup Yes For each tree, we compute its prediction performance before and after applying HS, where the regularization parameter for HS is selected from the set λ {0.1, 1.0, 10.0, 25.0, 50.0, 100.0} via cross-validation. Results for each experiment are averaged over 10 random data splits. In Fig S1, we simulate data via a linear model y = P10 i=1 xi + ϵ with x Unif[0, 1]50 and ϵ being drawn from a Gaussian or a Laplacian distribution for the left and right panel respectively, with noise variance σ2 = 0.01 in both cases. In both experiments, we used a training set of 500 samples to fit CART and hs CART models with a prescribed number of leaves, varying this number across a grid. For each hs CART model, the regularization parameter λ was chosen on the training set via 3-fold cross-validation. Finally, we repeat this entire process 100 times with resampled datasets.