Decomposing Global Feature Effects Based on Feature Interactions

Authors: Julia Herbinger, Marvin N. Wright, Thomas Nagler, Bernd Bischl, Giuseppe Casalicchio

JMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically evaluate the theoretical characteristics of the proposed methods based on various feature effect methods in different experimental settings. Moreover, we apply our introduced methodology to three real-world examples to showcase their usefulness.
Researcher Affiliation Academia 1 Department of Statistics, LMU Munich, Munich, Germany 2 Munich Center for Machine Learning (MCML), Munich, Germany 3 Leibniz Institute for Prevention Research and Epidemiology, Bremen, Germany 4 Faculty of Mathematics and Computer Science, University of Bremen, Bremen, Germany 5 Department of Public Health, University of Copenhagen, Copenhagen, Denmark
Pseudocode Yes Algorithm 1: Partitioning algorithm of GADGET; Algorithm 2: PINT
Open Source Code Yes All proposed methods and reproducible scripts for the experiments are available online via https://github.com/Julia Herbinger/gadget/.
Open Datasets Yes bikesharing data set (James et al., 2022); COMPAS data set... collected by Pro Publica (Larson et al., 2016); spam data set (Hopkins et al., 1999)
Dataset Splits Yes The R2 measured on a separately drawn test set of size 10000 following the same distribution is 0.94. (Section 4.3) ...measured by a 5-fold cross-validation, indicating good performance (Hopkins et al., 1999). (Section 9)
Hardware Specification No The paper describes various models (feed-forward neural network, GAM, XGBoost, random forest, SVM) and their configurations, but it does not specify the hardware (e.g., CPU, GPU models) used for training or evaluation.
Software Dependencies No The paper mentions 'R package version 1.3-2' for ISLR2 within a dataset citation, but this refers to a tool associated with a dataset, not the specific software dependencies with version numbers used for implementing the authors' methodology (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes We then draw 500 observations from these random variables and fit a feed-forward neural network (NN) with a single hidden layer of size 10 and weight decay of 0.001. (Section 4.3) As stopping criteria, we choose a maximum tree depth of 6, a minimum number of observations per leaf of 40, and set the improvement parameter γ to 0.2. (Section 6.1)