JevForest: Path Voting for Budgeted Feature Acquisition

arXiv:2610.10615v1 Announce Type: cross Abstract: Choosing which information to observe is central to prediction under limited observation budgets. We study JevForest, a feature acquisition policy that aggregates path-dependent proposals from bootstrapped trees, weights them by global training information gain, and predicts from the acquired values with a shared masked classifier. An online implementation queries Jev for semantic answers selected by this policy. On small balanced held-out samples, four-question forest acquisition achieves accuracy 0.729 on AG News (n=48), compared with 0.667 for a static gain ranking and 0.583 for random ordering. On TREC (n=24), the ordering reverses: forest accuracy is 0.667, compared with 0.750 and 0.833. Asking all eight questions in one batch yields higher accuracy at lower measured cost and latency than four sequential forest queries; direct Jev classification matches the batch accuracy while costing less. Offline MiniBooNE experiments yield accuracy 0.845pm0.010 at ten features and 0.885pm0.008 at forty features over three jointly varying data and forest seeds (mean pm sample standard deviation). A companion Newton boosting implementation provides preliminary full-feature synthetic results. These exploratory findings establish a working Jev acquisition workflow but do not support a general advantage for path voting: its value depends on the task, predictor, and the distinction between question budgets and actual query costs.

Source: arXiv cs.LG | 2026-10-09