Feature Responsiveness Scores: Model-Agnostic Explanations for Recourse

Authors: Seung Hyun Cheon, Anneke Wernerfelt, Sorelle Friedler, Berk Ustun

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct an extensive empirical study on the responsiveness of explanations in lending. Our results show that standard practices in consumer finance can backfire by presenting consumers with reasons without recourse, and demonstrate how our approach improves consumer protection by highlighting responsive features and identifying fixed predictions.
Researcher Affiliation Academia Seung Hyun Cheon UC San Diego Anneke Wernerfelt Haverford College Sorelle A. Friedler Haverford College Berk Ustun UC San Diego
Pseudocode Yes Algorithm 1 Sample Reachable Points Algorithm 2 Enumerate Reachable Points
Open Source Code Yes We include a Python library to compute feature responsiveness scores available on Git Hub.
Open Datasets Yes We work with three publicly available consumer finance classification datasets. ... heloc n = 5, 842 d = 43 FICO [23] ... german n = 1, 000 d = 36 Dua & Graff [15] ... givemecredit n = 120, 268 d = 23 Kaggle [32]
Dataset Splits Yes We split each dataset into a training sample (80%; to train models and tune parameters) and a test sample (20%; to evaluate out-of-sample performance).
Hardware Specification No The paper does not explicitly describe the hardware used to run its experiments.
Software Dependencies No The paper mentions using a 'Python library' and various machine learning models (Logistic Regression, XGBoost, Random Forests, SHAP, LIME) but does not provide specific version numbers for any of these software components.
Experiment Setup Yes We fit models using (1) logistic regression (LR), (2) XGBoost (XGB), and (3) random forests (RF). For each model, we construct featurehighlighting explanations for each person who is denied credit that highlight up to four features... We chose the sample size N = 500 to ensure that the 100(1 α)% confidence interval in Appendix A.2 had an upper bound 0.01 when ˆµj(x) = 0 with α = 0.01.