JevForest: Path Voting for Budgeted Feature Acquisition

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 0.845±0.010 at ten features and 0.885±0.008 at forty features over three jointly varying data and forest seeds (mean ± 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.