SafeAR: Safe Algorithmic Recourse by Risk-Aware Policies

Authors: Haochen Wu, Shubham Sharma, Sunandita Patra, Sriram Gopalakrishnan

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We apply our method to two real-world datasets and compare policies with different risk-aversion levels using risk measures and recourse desiderata (sparsity and proximity). Evaluate the policies with different risk profiles computed by G-RSVI on two real-world datasets (UCI Adult Income, German Credit)
Researcher Affiliation Collaboration 1 University of Michigan, Ann Arbor 2 J.P. Morgan AI Research haochenw@umich.edu, shubham.x2.sharma@jpmchase.com, sunandita.patra@jpmchase.com, sriram.gopalakrishnan@jpmchase.com
Pseudocode Yes Algorithm 1: G-RSVI Input: recourse MDP S, A, T, R, H , ML model f Parameters: risk aversion level β [0, ]
Open Source Code Yes All supplemental materials (appendices and code implementations) are available through arxiv.org/abs/2308.12367.
Open Datasets Yes Adult Income Dataset (AID) (32561 data points) (Becker and Kohavi 1996) and German Credit Dataset (GCD) (1000 data points) (Hofmann 1994)
Dataset Splits No The paper mentions converting continuous features and training classifiers, but does not explicitly detail train/validation/test splits with percentages, sample counts, or specific methodologies for reproducibility. It states 'All measures are averaged over the entire dataset'.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU, GPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions training 'random forest classifiers' but does not specify any software names with version numbers (e.g., scikit-learn version, Python version, or other libraries).
Experiment Setup Yes The horizon is set to 12. We select β = 0.25, 0.50, 0.75 for generating risk-averse recourse policies, and higher β indicates higher risk-aversion. We use qualitative assumptions (domain knowledge) on relative differences in action costs and success likelihood to define the action costs r( ) and transition model p( ). The transition probabilities are heuristically set by domain knowledge.