[論文レビュー] 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 AI被引用数 43

ひとこと要約

この論文は、対照的な説明と生じる推奨に関するアルゴリズム的救済の定義、定式化、解法を統合し、今後の研究方向を概説します。

ABSTRACT

Machine learning is increasingly used to inform decision-making in sensitive\nsituations where decisions have consequential effects on individuals' lives. In\nthese settings, in addition to requiring models to be accurate and robust,\nsocially relevant values such as fairness, privacy, accountability, and\nexplainability play an important role for the adoption and impact of said\ntechnologies. In this work, we focus on algorithmic recourse, which is\nconcerned with providing explanations and recommendations to individuals who\nare unfavourably treated by automated decision-making systems. We first perform\nan extensive literature review, and align the efforts of many authors by\npresenting unified definitions, formulations, and solutions to recourse. Then,\nwe provide an overview of the prospective research directions towards which the\ncommunity may engage, challenging existing assumptions and making explicit\nconnections to other ethical challenges such as security, privacy, and\nfairness.\n

研究の動機と目的

• Synthesize and unify existing definitions of algorithmic recourse (contrastive explanations vs consequential recommendations).
• Present unified formulations and solution strategies for computing recourse under realism constraints (actions, plausibility, and causality).
• Map the landscape of technical approaches across model types and data settings to guide future research and practice.

提案手法

• Define contrastive explanations as nearest counterfactuals to a fixed model using a plausible data subspace.
• Define consequential recommendations as interventions in a structural causal model to achieve favorable outcomes with minimal cost.
• Present constrained optimization formulations for both explanations (Equation 1) and recommendations (Equation 2).
• Discuss dissimilarity (dist) and cost measures (e.g., Manhattan distance, weighted norms) and their trade-offs.
• Characterize actionability, plausibility, and diversity/sparsity constraints and their role in recourse generation.
• Outline data types and encoding practices for tabular, image, and text data.

実験結果

リサーチクエスチョン

• RQ1 What are the unified definitions of recourse that cover both explanations and recommendations?
• RQ2 How can recourse be formulated and solved under causal, actionable, and plausibility constraints?
• RQ3 What are the trade-offs between nearest contrastive explanations and minimal-cost consequential recommendations?
• RQ4 How do data types and encodings affect the generation of recourse?
• RQ5 What future directions and ethical considerations shape the trajectory of algorithmic recourse research?

主な発見

• Most literature focuses on contrastive explanations rather than consequential recommendations.
• A causal-model perspective enables actionable, minimal-cost recommendations with downstream effects accounted for, beyond simple feature shifts.
• Recourse formulations depend on dist and cost metrics, with trade-offs between sparsity, diversity, and plausibility.
• Actionability and plausibility constraints are distinct but complementary in shaping feasible recourse.
• A broad taxonomy covers model types (tree-based, kernel-based, differentiable) and data types (tabular, images, text).
• The paper identifies future research directions linking recourse to fairness, privacy, and security in ethical ML.