Towards User Guided Actionable Recourse

User preferences matter for recourse recommendations: The paper identifies that existing AI recourse tools focus on technical metrics like proximity and cost, but overlook what users actually want to change about themselves. By incorporating user preferences through soft constraints, the approach generates recommendations that are more likely to be followed in practice.

Three practical ways to capture what users prefer: The paper proposes simple, implementable methods for understanding user preferences: scoring how much they value changing different features, setting acceptable ranges for feature values, and ranking which categorical options they prefer—making the system accessible to non-technical users.

Actionable recourse is critical in high-stakes domains: The paper emphasizes that when ML systems deny loans, job opportunities, or parole, users need clear, feasible guidance on what changes could lead to better outcomes. Without considering user preferences, recommended changes may be unrealistic or unacceptable, rendering the recourse useless.

Current recourse systems may give impractical advice: Existing approaches optimize for mathematical properties but can recommend changes users cannot or will not make. By centering user preferences, the proposed User Preferred Actionable Recourse (UP-AR) approach aims to bridge the gap between what's technically optimal and what's actually achievable for real people.