[Paper Review] Algorithmic Recourse: from Counterfactual Explanations to Interventions

Amir-Hossein Karimi, Bernhard Schölkopf+1 more|arXiv (Cornell University)|Feb 14, 2020

Explainable Artificial Intelligence (XAI)Computer Science50 references72 citations

TL;DR

The paper argues that counterfactual explanations alone are insufficient for actionable recourse and introduces a minimal intervention framework (MINT) using causal models to provide cost-efficient, feasible actions that change predictions. It shows that recourse via minimal interventions can outperform traditional counterfactual-explanation-based recourse in both synthetic and real-world settings.

ABSTRACT

As machine learning is increasingly used to inform consequential decision-making (e.g., pre-trial bail and loan approval), it becomes important to explain how the system arrived at its decision, and also suggest actions to achieve a favorable decision. Counterfactual explanations -- "how the world would have (had) to be different for a desirable outcome to occur" -- aim to satisfy these criteria. Existing works have primarily focused on designing algorithms to obtain counterfactual explanations for a wide range of settings. However, one of the main objectives of "explanations as a means to help a data-subject act rather than merely understand" has been overlooked. In layman's terms, counterfactual explanations inform an individual where they need to get to, but not how to get there. In this work, we rely on causal reasoning to caution against the use of counterfactual explanations as a recommendable set of actions for recourse. Instead, we propose a shift of paradigm from recourse via nearest counterfactual explanations to recourse through minimal interventions, moving the focus from explanations to recommendations. Finally, we provide the reader with an extensive discussion on how to realistically achieve recourse beyond structural interventions.

Motivation & Objective

• Motivate the need for actionable recourse beyond counterfactual explanations in consequential decisions.
• Critique the assumption that nearest counterfactuals directly translate to feasible actions for individuals.
• Propose a causal reformulation that uses minimal interventions to achieve favorable model outcomes.
• Demonstrate, via synthetic and real-world data, that minimal interventions can be cheaper and more feasible than CFE-based recourse.

Proposed method

• Analyze limitations of existing counterfactual explanations by embedding them in a structural causal model (SCM).
• Formalize actions as interventions using the do-operator and compute structural counterfactuals via Abduction-Action-Prediction.
• Define CFE-based actions and show necessary/sufficient conditions for guaranteed recourse under causal dependencies.
• Introduce and solve the Minimal Interventions (MINT) optimization that minimizes action cost while ensuring h(xSCF) changes.

Experimental results

Research questions

• RQ1 What are the intrinsic limitations of deriving recourse actions directly from counterfactual explanations under causal dependencies?
• RQ2 Can recourse be achieved more reliably and cheaply by framing actions as minimal interventions within a causal model?
• RQ3 How does the proposed Minimal Interventions framework compare to CFE-based recourse in both synthetic and real-world settings?

Key findings

• CFE-based recourse can be suboptimal or infeasible in non-independent worlds due to causal dependencies.
• Recasting recourse as minimal interventions yields guaranteed recourse at lower cost when feasible actions are chosen over raw counterfactual shifts.
• In synthetic and German credit dataset experiments, minimal interventions require less effort for individuals to achieve favorable model outcomes compared to CFE-based actions.
• The Abduction-Action-Prediction framework enables computing structural counterfactuals for any feasible action set within additive noise SCMs.
• Minimal Interventions (MINT) provides a principled way to select actions that account for downstream effects of interventions on non-intervened variables.