Sriram Gopalakrishnan, Ph.D.

Sriram Gopalakrishnan, Ph.D.

AI Research Lead at J.P. Morgan Chase | Sequential-Decision Making Problems, Human-AI Interaction, & Generative AI

SafeAR: Risk-Aware Algorithmic Recourse

Published:

Pillar: Human-AI Interaction
Algorithmic Recourse CVaR Human-in-the-Loop Safe RL

Overview

Algorithmic Recourse empowers users to reverse unfavorable algorithmic decisions (e.g. loan denials or credit limits). However, classical recourse algorithms implicitly assume that whenever a user attempts an action, it succeeds deterministically. In reality, human efforts have uncertain outcomes.

SafeAR reformulates algorithmic recourse through the lens of risk-sensitive Markov Decision Processes:

  • Minimizes the probability of catastrophic downside trajectories using Conditional Value-at-Risk (CVaR) criteria.
  • Provides actionable, robust recourse policies that remain safe even under significant human execution noise.
  • Recognized with an Oral Presentation (Top 2%) at AAAI 2024.