As machine‑learning systems increasingly influence individual lives, providing users with actionable recourse becomes crucial. Existing solutions rely on costly human supervision, requiring a hand‑crafted cost function before searching for low‑cost actions. This paper introduces Data‑Driven Action Recourse (DDAR), the first method that delivers recourse without any human‑annotated action costs. DDAR rests on the observation that historically frequent user trajectories should correspond to lower costs. Accordingly, DDAR first learns a dynamics model from past user trajectories, then searches within this model for trajectories that achieve the desired outcome. Experiments show that DDAR produces plausible recourse suggestions without human supervision and achieves significantly lower average action costs than prior approaches.
Core Techniques
- Dynamics model: learn transition probabilities $P(s_{t+1}\mid s_t, a_t)$ from historical sequences.
- Trajectory search: find an action sequence minimizing $\sum_{t} c(a_t)$ while satisfying the target prediction under the learned model.
- Cost estimation: approximate costs by the inverse frequency of observed trajectories, eliminating the need for manual labeling.
Experimental Findings Across several benchmark datasets, DDAR’s recommended actions are more feasible and reduce average cost by roughly 15% compared to baselines.
Review