Shapley Residuals: Quantifying the limits of the Shapley value for explanations

Authors: Indra Kumar, Carlos Scheidegger, Suresh Venkatasubramanian, Sorelle Friedler

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In summary, we: ... show via a number of experiments that Shapley residuals capture meaningful information for model explanations in realistic scenarios (Section 6) and Having theoretically justified Shapley residuals in previous sections, we now focus on illustrating what these residuals can help us understand about models on a real-world dataset.
Researcher Affiliation Academia I. Elizabeth Kumar Department of Computer Science Brown University Providence, RI 02912 iekumar@brown.edu Carlos Scheidegger Department of Computer Science University of Arizona Tucson, AZ 85721 cscheid@cs.arizona.edu Suresh Venkatasubramanian Department of Computer Science Brown University Providence, RI 02912 suresh@brown.edu Sorelle A. Friedler Computer Science Dept. Haverford College Haverford, PA 19041 sorelle@cs.haverford.edu
Pseudocode Yes Algorithm 1 Exactly calculate the ith Shapley value and Shapley residual of v
Open Source Code Yes Code is provided in the supplementary material
Open Datasets Yes Consider the Shapley values and residuals for an occupancy detection dataset3 (20,560 instances) used to predict whether an office room is occupied. 3https://archive.ics.uci.edu/ml/datasets/Occupancy+Detection+
Dataset Splits Yes A decision tree model with maximum depth 3 is trained on 75% of the data using the features light and hour.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU types, or memory specifications used for running experiments.
Software Dependencies No The paper mentions using "our own implementation of Kernel SHAP" but does not specify any software dependencies with version numbers (e.g., Python, PyTorch, specific library versions).
Experiment Setup Yes A decision tree model with maximum depth 3 is trained on 75% of the data using the features light and hour. and We then calculate the Shapley values and residuals (using 50 randomly sampled background rows from the test set) for 1000 randomly sampled test instances.