[Paper Review] Towards Robust and Reliable Algorithmic Recourse

Sohini Upadhyay, Shalmali Joshi+1 more|arXiv (Cornell University)|Feb 26, 2021

Explainable Artificial Intelligence (XAI)Computer Science31 references37 citations

TL;DR

ROAR proposes a minimax framework to generate counterfactual recourses that remain valid under model shifts, with theoretical bounds and strong empirical robustness (67–100% more robust than baselines) across real-world datasets.

ABSTRACT

As predictive models are increasingly being deployed in high-stakes decision making (e.g., loan approvals), there has been growing interest in post hoc techniques which provide recourse to affected individuals. These techniques generate recourses under the assumption that the underlying predictive model does not change. However, in practice, models are often regularly updated for a variety of reasons (e.g., dataset shifts), thereby rendering previously prescribed recourses ineffective. To address this problem, we propose a novel framework, RObust Algorithmic Recourse (ROAR), that leverages adversarial training for finding recourses that are robust to model shifts. To the best of our knowledge, this work proposes the first solution to this critical problem. We also carry out detailed theoretical analysis which underscores the importance of constructing recourses that are robust to model shifts: 1) we derive a lower bound on the probability of invalidation of recourses generated by existing approaches which are not robust to model shifts. 2) we prove that the additional cost incurred due to the robust recourses output by our framework is bounded. Experimental evaluation on multiple synthetic and real-world datasets demonstrates the efficacy of the proposed framework and supports our theoretical findings.

Motivation & Objective

• Motivate the need for recourse methods that remain valid when predictive models are updated due to data distribution shifts.
• Propose a minimax objective to produce robust counterfactuals that are resistant to plausible model shifts.
• Provide theoretical bounds on the probability of invalidation for non-robust recourses and on the cost increase for robust recourses.
• Demonstrate empirical robustness of ROAR across synthetic and real-world datasets under various shift scenarios.

Proposed method

• Introduce ROAR, a minimax objective that minimizes the worst-case loss over a set of plausible model shifts Delta.
• Approximate general predictive models with a local linear surrogate f_w to enable perturbations delta in the parameter space.
• Define the counterfactual recourse x'' as the solution to min_{x'' in A} max_{delta in Delta} ell(f_{w+delta}(x''), 1) + lambda c(x, x'').
• Specify Delta using either per-feature bounds (delta_min <= delta_i <= delta_max) or a norm ball (||delta||_p <= delta_max).
• Develop an adversarial-training-inspired optimization procedure (Algorithm 1) that alternates maximizing the model perturbation and updating the recourse via gradients until convergence.

Experimental results

Research questions

• RQ1 How vulnerable are existing recourse methods to model shifts induced by distribution changes?
• RQ2 Can we formulate and optimize recourses that remain valid under plausible model perturbations while controlling recourse cost?
• RQ3 What are the theoretical limits (lower/upper bounds) on invalidation probability and cost inflation when using robust recourses?
• RQ4 Do robust recourses improve real-world robustness across datasets with correction, temporal, and geospatial shifts without excessive cost?

Key findings

• ROAR-based recourses are substantially more robust to model shifts (67–100% increase in robustness) than state-of-the-art baselines across real-world datasets.
• ROAR achieves high validity on the original model (often >95%) while significantly improving robustness to shifted models.
• The framework provides theoretical guarantees: a lower bound on the invalidation probability for non-robust recourses and an upper bound on the additional cost of robust recourses.
• Robust recourses incur bounded cost increases relative to non-robust ones under modeled shifts, preserving practical feasibility.
• Empirical evaluation on German credit, SBA, and geospatial/temporal datasets demonstrates ROAR’s effectiveness over multiple baselines (CFE, AR, MINT, ROAR variants).