JevForest
An arXiv paper addressing the budget-constrained feature acquisition problem. It proposes a feature acquisition strategy based on bootstrap decision tree path voting, using globally trained information-gain weighting and a shared mask classifier to achieve dynamic feature selection under cost constraints.
Timeline (last 90 days)
JevForest is a strategy proposed for the budget-constrained feature acquisition problem, based on bootstrap-sampled decision trees performing path voting.
Each tree in JevForest proposes which feature should be observed next along its decision path, making it a path-dependent proposal approach.
JevForest weights the proposals from different trees according to global training information gain, so features with higher information gain receive greater weight in voting.
JevForest uses a shared masked classifier to make predictions based on already-observed feature values, where the masking mechanism enables the model to handle partially missing inputs.
On the AG News dataset (n=48), using a forest acquisition approach with four questions achieved an accuracy of 0.729, outperforming static gain ranking at 0.667 and random ordering at 0.583.
Asking all eight questions at once in a single batch not only achieves higher accuracy, but also exhibits lower measured cost and latency than four sequential forest queries.
In offline MiniBooNE experiments, JevForest achieved an accuracy of 0.845±0.010 using ten features, improving to 0.885±0.008 using forty features.
The paper provides an accompanying Newton boosting implementation and presents preliminary full-feature synthetic data results.
The paper's authors position these results as exploratory, explicitly stating that the existing evidence does not support a universal advantage for path voting.
The experimental results of JevForest exhibit clear task dependence, with diametrically opposite performance on AG News and TREC, two similar text classification tasks.
3 more timeline events
All Facts (13)
JevForest is a strategy proposed for the budget-constrained feature acquisition problem, based on bootstrap-sampled decision trees performing path voting.
50%UnverifiedEach tree in JevForest proposes which feature should be observed next along its decision path, making it a path-dependent proposal approach.
50%UnverifiedJevForest weights the proposals from different trees according to global training information gain, so features with higher information gain receive greater weight in voting.
50%UnverifiedJevForest uses a shared masked classifier to make predictions based on already-observed feature values, where the masking mechanism enables the model to handle partially missing inputs.
50%UnverifiedOn the AG News dataset (n=48), using a forest acquisition approach with four questions achieved an accuracy of 0.729, outperforming static gain ranking at 0.667 and random ordering at 0.583.
50%UnverifiedAsking all eight questions at once in a single batch not only achieves higher accuracy, but also exhibits lower measured cost and latency than four sequential forest queries.
50%UnverifiedIn offline MiniBooNE experiments, JevForest achieved an accuracy of 0.845±0.010 using ten features, improving to 0.885±0.008 using forty features.
50%UnverifiedThe paper provides an accompanying Newton boosting implementation and presents preliminary full-feature synthetic data results.
50%UnverifiedThe paper's authors position these results as exploratory, explicitly stating that the existing evidence does not support a universal advantage for path voting.
50%UnverifiedThe experimental results of JevForest exhibit clear task dependence, with diametrically opposite performance on AG News and TREC, two similar text classification tasks.
50%UnverifiedThe question budget and actual query cost are two different things; sequential decision-making is more economical in terms of number of questions, but may not be cost-effective in terms of real cost and latency.
50%UnverifiedBefore adopting any intelligent feature acquisition strategy, thorough validation on the specific task should be performed, rather than assuming that dynamic path voting is inherently superior to static ranking or batch querying.
50%UnverifiedThe arXiv number of the JevForest paper is arXiv:2610.10615v1.
50%