Adapting and Evaluating Influence-Estimation Methods for Gradient-Boosted Decision Trees

Authors: Jonathan Brophy, Zayd Hammoudeh, Daniel Lowd

JMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments on 22 real-world data sets to compare eight different influence-estimation methods using 5 different evaluation measures and 4 popular modern GBDT implementations.
Researcher Affiliation Academia Jonathan Brophy EMAIL Zayd Hammoudeh EMAIL Daniel Lowd EMAIL Department of Computer and Information Science University of Oregon Eugene, OR 97403, USA
Pseudocode No The paper describes methods with mathematical formulations and prose, but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes source code is available at https://github.com/jjbrophy47/tree_influence.
Open Datasets Yes We evaluate on 22 real-world tabular data sets (13 binary-classification tasks, 1 multiclass-classification task, and 8 regression tasks) well-suited for boosted tree-based models... All additional data set details are in the Appendix, A.1. ... Adult (Dua and Graff, 2019) contains 48,842 instances... Bank (Moro et al., 2014; Dua and Graff, 2019) consists of 41,188 marketing phone calls...
Dataset Splits Yes For any data set without a predefined train/test split, we sample 80% of the data uniformly at random for training and use the rest for testing... Each model is tuned using 5-fold cross-validation
Hardware Specification Yes Experiments are run on an Intel(R) Xeon(R) CPU E5-2690 v4 @ 2.6GHz with 100GB of DDR4 RAM @ 2.4GHZ.
Software Dependencies Yes Experiments are run using Python 3.9.6, and source code for all influence-estimation implementations and all experiments is available at https://github.com/jjbrophy47/tree_influence.
Experiment Setup Yes Each model is tuned using 5-fold cross-validation; selected hyperparameters are in A.2: Table 4, with predictive performance comparisons in A.2: Table 3. ... For the LGB, XGB, CB, and SGB models, we tune the number of trees/boosting iterations (T) using values [10, 25, 50, 100, 200]. Since the LGB and SGB models grow trees in a leaf-wise (depth-first) manner, we tune the maximum number of leaves (lmax) for LGB and SGB using values [15, 31, 61, 91]. In contrast, we tune the the maximum depth (dmax) for XGB and CB using values [2, 3, 4, 5, 6, 7]. We also tune the learning rate (η) for CB using values [0.1, 0.3, 0.6, 0.9], and the maximum number of bins (bmax) for SGB using values [50, 100, 250]. We tune all hyperparameters using 5-fold cross-validation.