[Paper Review] A survey of algorithmic recourse: definitions, formulations, solutions, and prospects

Amir-Hossein Karimi, Gilles Barthe+2 more|arXiv (Cornell University)|Oct 8, 2020

Ethics and Social Impacts of AI197 references45 citations

TL;DR

This paper surveys algorithmic recourse, distinguishing contrastive explanations from consequential recommendations, and provides unified definitions, formulations, and solutions, while outlining future research directions and ethical connections.

ABSTRACT

Machine learning is increasingly used to inform decision-making in sensitive situations where decisions have consequential effects on individuals' lives. In these settings, in addition to requiring models to be accurate and robust, socially relevant values such as fairness, privacy, accountability, and explainability play an important role for the adoption and impact of said technologies. In this work, we focus on algorithmic recourse, which is concerned with providing explanations and recommendations to individuals who are unfavourably treated by automated decision-making systems. We first perform an extensive literature review, and align the efforts of many authors by presenting unified definitions, formulations, and solutions to recourse. Then, we provide an overview of the prospective research directions towards which the community may engage, challenging existing assumptions and making explicit connections to other ethical challenges such as security, privacy, and fairness.

Motivation & Objective

• Unify definitions, formulations, and technical solutions for recourse across a broad range of setups.
• Differentiate contrastive explanations from consequential recommendations within causal frameworks.
• Summarize constraints and practical considerations such as actionability, plausibility, diversity, and sparsity.
• Position recourse within broader ethical ML topics like fairness, privacy, and accountability.

Proposed method

• Present unified definitions distinguishing contrastive explanations and consequential recommendations (Q1 vs Q2).
• Formulate recourse as constrained optimization problems with equations (1) and (2) for counterfactuals and actions.
• Discuss dist and cost metrics (e.g., Manhattan distance with MAD, mixed ||d||p norms) and how they shape solutions.
• Categorize model types (tree-based, kernel-based, differentiable, and other) and data types used in recourse.
• Detail actionability and plausibility constraints, and how diversity and sparsity are incorporated into solutions.
• Provide a survey of 50+ technical papers (Table 1) and highlight open challenges and future directions.

Experimental results

Research questions

• RQ1 What constitutes recourse in algorithmic decision-making, and how can it be unified across explanations and recommendations?
• RQ2 How can we formulate and solve for contrastive explanations and consequential recommendations under realistic constraints?
• RQ3 What are the roles of causality, actionability, and plausibility in generating feasible recourse?
• RQ4 What are the main methodological gaps and future directions for integrating recourse with broader ethical ML concerns?

Key findings

• Most recourse work focuses on generating contrastive explanations rather than consequential recommendations.
• Differentiable models are the most widely supported class in recourse methods.
• Solutions trade off properties like optimality, coverage, runtime, and access, with no single unifying benchmark.
• Consequential recommendations rely on causal models (SCMs) and may be computationally more demanding due to abduction-action-prediction steps.
• Actionability and plausibility constraints are distinct and must be jointly considered to yield feasible and believable recourse.
• The survey highlights a broad set of papers (Table 1) and identifies future research directions linking recourse to security, privacy, and fairness.