Learning Source Acquisition Policies by Offline Planning

Summary

The paper proposes O-MPAC, a method for budgeted active feature acquisition in which an external schema defines feature-group sources with costs. An offline teacher evaluates finite-horizon (one- or two-step) risk-cost backups on complete training records using a fixed predictor, producing per-source advantage targets relative to stopping. A shared source-action scorer, trained on partial observations and source metadata, imitates these value, ranking, and optionally action targets, then re-scores all affordable sources after every query under a hard cost mask. The analysis covers action-target decomposition, permutation-invariant uniform supervision over tied minima, and exact routing-target gaps. In CUBE-NM routing controls, uniform tie supervision reaches 0.965 accuracy under both original and context-last source orders, whereas first-index action labels collapse to 0.295 when context moves last; matching the remaining horizon separates useful from redundant final queries. On six real tabular tasks with five seeded splits, validation selects the H1 noCE configuration in all thirty splits, and O-MPAC attains the highest mean budget-integrated accuracy (nAUPC) on five of six tasks against adapted GDFS, DIME, AACO+NN and a static policy (0.760 versus 0.735-0.746); fourteen of twenty-four paired comparisons pass BH correction, five pass Bonferroni. Limitations include small real-task margins, H2 never selected on real data, a weak centroid-based teacher, synthetic costs, and no stated code release.