A survey of algorithmic recourse: definitions, formulations, solutions, and prospects (2010.04050v2)

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.

• The paper synthesizes current literature by distinguishing contrastive explanations from consequential recommendations to empower algorithmic recourse.
• It formulates recourse generation as a constrained optimization problem, categorizing solutions by model type and action feasibility.
• Its comprehensive review identifies trade-offs and proposes future interdisciplinary research to enhance transparency and user agency in ML systems.

A Survey of Algorithmic Recourse: Contrastive Explanations and Consequential Recommendations

The paper "A Survey of Algorithmic Recourse: Contrastive Explanations and Consequential Recommendations" by Karimi et al. tackles a critical aspect of the interface between ML systems and human users, specifically addressing the need for recourse in decision-making systems. Algorithmic recourse refers to the ability of individuals to understand and potentially change unfavorable decisions rendered by ML models that increasingly permeate sensitive domains such as finance, healthcare, and justice.

Content Overview

Algorithmic recourse is defined in the paper through the dual concepts of contrastive explanations and consequential recommendations. The work begins with a clear distinction between these two notions. Contrastive explanations provide insights into why a particular decision was made over another, positing alternative scenarios that could have led to a different decision. In contrast, consequential recommendations are actionable advice that individuals can undertake to achieve a favorable outcome in the future. This distinction is crucial for understanding the different levels of intervention required in the causal history of decision-making processes.

The authors undertake a comprehensive survey of the existing literature on algorithmic recourse up to the year 2020, providing a valuable consolidation of definitions, formulations, and solutions concerning recourse. They emphasize the importance of understanding causal relationships inherent in decision-making models, illustrating how these impact the provision of explanations and recommendations.

Methodological Contributions

The paper effectively organizes algorithmic recourse solutions into a methodological framework that considers various model types (e.g., tree-based, kernel-based, differentiable), alongside actionability and plausibility constraints. This categorization aids in managing the complexity inherent in designing systems that can provide both meaningful explanations and viable recommendations.

A significant methodological contribution is the formulation of recourse generation as a constrained optimization problem. Here, the aim is to define objective functions that measure dissimilarity between feature sets in explanations and the cost of actions in recommendations. This framework is used to evaluate the feasibility of obtaining specific changes through actionable interventions, highlighting the technical considerations necessary for different ML model architectures.

Numerical Results and Implications

While the paper does not focus on specific numerical results, it provides an extensive tabular summary of over 50 influential papers in the recourse literature. This summary categorizes each approach by its objectives, constraints, data types, and other salient properties. The researchers systematically identify trade-offs and challenges in achieving certain desirable properties, such as optimization, coverage, and runtime efficiency.

The discussion extends to speculate on future trends and challenges in the field, proposing directions for further research that addressed limitations in current approaches. The authors stress the importance of interdisciplinary collaboration to evolve recourse solutions that are both technically robust and socially considerate.

Theoretical and Practical Implications

The paper's insights generate profound implications for both theoretical and practical perspectives in AI. Theoretically, it challenges existing assumptions about the causal nature of decision-making processes in ML, positing a need for robust frameworks that integrate counterfactual reasoning. Practically, it underscores the potential for algorithmic recourse to enhance trust and transparency in ML systems. The notion that recourse recommendations demand more than the identification of plausible changes but also necessitate actionable pathways highlights a crucial intersection of ML design and human agency.

In conclusion, this survey represents a foundational effort in synthesizing the current state of algorithmic recourse. By proposing a unified view of the field and identifying practical and theoretical challenges, it sets the stage for continued advancement in equitably integrating ML decision-making systems within ethically nuanced societal frameworks.