MAPLE: Multi-Agent Prior Learning for Constructing Tree Ensembles

Nguyen Viet Tuan Kiet, Nguyen Ba Thinh, Thanh Trung Huynh, Hieu Pham

Keywords: Multi-Agent Learning, Tabular Learning, Ensemble Learning, Decision Tree

TL;DR: MAPLE integrates LLM-derived feature priors into tree ensemble construction via multi-agent learning, improving performance and robustness on tabular data at scale.

Abstract: Tree ensembles based on bagging and boosting remain highly effective for tabular data, yet their construction typically relies on uniform or heuristic feature sampling strategies that overlook task-specific prior knowledge. We introduce MAPLE, a framework for Multi-Agent Prior Learning that integrates learned feature priors directly into the process of constructing tree ensembles. By leveraging multiple sources of inductive bias, MAPLE enables the ensemble to incorporate informative priors while preserving diversity. Experiments on multiple tabular benchmarks demonstrate that MAPLE consistently improves predictive performance and robustness over standard tree ensembles and prior-agnostic baselines, while remaining scalable and computationally efficient.

Submission Number: 56