Reproduce “Bagged Polynomial Regression and Neural Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sylvia Klosin, Jaume Vives-i-Bastida
Researchers at UC Davis and Stanford propose Bagged Polynomial Regression (BPR) and Partitioned Polynomial Regression (PPR) as transparent alternatives to neural networks. BPR achieves competitive predictive accuracy in high-dimensional environmental applications while offering superior interpretability through coefficient-based structures.
Key insights
- Covariate partitioning in series estimators can significantly improve L2 convergence rates by controlling the growth of the dictionary basis, effectively mitigating the curse of dimensionality for smooth target functions.
- An optimal partitioning strategy exists that minimizes finite-sample L2 risk, with the optimal group size increasing with polynomial degree and smoothness.
- Models exhibiting sparsely supported derivative basis functions, like BPR, can yield faster and more reliable estimation error rates for marginal effects, enhancing the interpretability of how individual features impact predictions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.