Explanations-Driven Active Feature Acquisition for Algorithmic Recourse

September 14, 2026 | Source: arXiv AI

Tags: explainability, algorithmic-recourse, feature-acquisition, Markov-Blanket, XAI, fairness

EDFA uses Markov Blanket theory to unify counterfactual and semifactual explanations, selecting features by explanatory value per unit cost for cost-constrained prediction. On 7 public datasets, it acquires substantially fewer features than state-of-the-art baselines while maintaining accuracy and providing distribution-free recourse validity guarantees.

Details

Galwaduge and Samarabandu address the gap between algorithmic recourse (explaining what to change to flip a prediction) and active feature acquisition (deciding which features to measure given costs). Existing recourse methods assume all features are known; AFA methods ignore explanatory value. EDFA combines both. The framework uses Markov Blanket theory to unify counterfactual, semifactual, and alterfactual explanations under a common formalism and characterize how available recourse grows as features are acquired. EDFA selects the next feature to acquire based on explanatory value per unit cost — not just predictive accuracy per cost. A key practical contribution is distribution-free validity guarantees for recourse issued from partial feature information. These guarantees signal when lower-cost, partial-information recourse is trustworthy, along with a lower bound on the calibration data required to certify them. Practitioners can know before acquiring expensive features whether existing features support trustworthy recourse. Experiments on 7 publicly available datasets with neural network predictors show EDFA acquires substantially fewer features than state-of-the-art AFA baselines while maintaining comparable accuracy and producing more decision-relevant, actionable recourse. Implementation is available on GitHub.