Human-in-Loop AI Provides Personalized Causal Recourse.

When machine learning models make unfavorable decisions, providing users with actionable recommendations, known as algorithmic recourse, is crucial. Traditional recourse methods often fall short by offering generic counterfactual explanations or assuming a pre-defined understanding of a user's underlying causal structure. This can lead to interventions that are not truly personalized or relevant to an individual's unique context. This research proposes a novel human-in-the-loop framework designed to overcome these limitations. The framework iteratively refines its understanding of a user's structural causal model by engaging in interactive queries and employing Bayesian inference. This dynamic interaction allows the system to learn and approximate the specific causal dependencies relevant to each user. By incorporating human feedback, the framework improves its ability to identify causal effects, leading to recourse recommendations that are more plausible, cost-effective, and genuinely aligned with the user's actual circumstances. While simulations show promising results for both linear and non-linear causal models, the authors note that capturing highly complex, non-linear structures remains a challenge, underscoring the importance of accurate approximations and robust noise modeling.