Individual and Group-level considerations of Actionable Recourse
Jayanth Yetukuri · 2023
The advent of machine learning in several critical fields, such as banking, healthcare, and criminal justice, has inspired research into improving robustness, trustworthiness, and transparency in the models. Actionable Recourse is one such tool that enables the negatively impacted users to receive a favorable outcome by providing recommendations of cost-efficient changes to their features. Current recourse methodologies optimize for proximity, sparsity, validity, and distance-based costs. Actionability takes both individual and group-level signals. A critical component of actionability is the consideration of User Preference to guide the recourse generation process. These preferences can take several forms, and we introduce three such preferences to capture the individual difficulty of user actions. Additionally, feasibility and plausibility should be considered as a fixed set of pre-specified constraints. We argue that plausibility draws strong signals from group-level population information, which must be considered to achieve low-cost recourses across protected groups. Recoursability is an active research area, and plausibility becomes an essential direction for further research.