Fast Sparse Decision Tree Optimization via Reference Ensembles
Authors: Hayden McTavish, Chudi Zhong, Reto Achermann, Ilias Karimalis, Jacques Chen, Cynthia Rudin, Margo Seltzer9604-9613
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments show that in many cases we can rapidly construct sparse decision trees that match the accuracy of black box models. |
| Researcher Affiliation | Academia | Hayden Mc Tavish1,3*, Chudi Zhong2*, Reto Achermann1, Ilias Karimalis1, Jacques Chen1, Cynthia Rudin2, Margo Seltzer1 1 University of British Columbia 2 Duke University 3 University of California, San Diego |
| Pseudocode | No | The paper describes steps for an algorithm in section 4.3 using numbered points, but it is presented as descriptive text rather than structured pseudocode or an algorithm block. |
| Open Source Code | Yes | Code Availability Implementations of GOSDT and DL8.5 with the guessing strategies discussed in this paper are available at https://github.com/ubc-systopia/gosdt-guesses and https://github.com/ubc-systopia/pydl8.5-lbguess. |
| Open Datasets | Yes | We use seven datasets: one simulated 2D spiral pattern dataset, the Fair Isaac (FICO) credit risk dataset (FICO et al. 2018) for the Explainable ML Challenge, three recidivism datasets (COMPAS, Larson et al. 2016, Broward, Wang et al. 2022, Netherlands, Tollenaar and Van der Heijden 2013), and two coupon datasets (Takeaway and Restaurant), which were collected on Amazon Mechanical Turk via a survey (Wang et al. 2017). |
| Dataset Splits | Yes | Unless stated otherwise, all plots show the median value across 5 folds with error bars corresponding to the first and third quartile. |
| Hardware Specification | No | We limit run time to 30 minutes and run on a machine with 125GB memory. |
| Software Dependencies | No | We use GBDT as the reference model for guessing and run it using scikit-learn (Pedregosa et al. 2011). |
| Experiment Setup | Yes | We configure GBDT with default parameters, but select dataset-specific values for the depth and number of weak classifiers: (n_est, max_depth) = (20, 3) for COMPAS and Spiral, (40, 1) for Broward and FICO, (50, 2) for Takeaway and Restaurant, and (30, 2) for Netherlands. Appendix C presents details about our hyper-parameter selection process, and Appendix B presents the experimental setup and more details about the datasets. |