A0328
Title: SplitWise regression: Stepwise modeling with adaptive dummy encoding
Authors: Marcell Tamas Kurbucz - University College London (United Kingdom) [presenting]
Nikolaos Tzivanakis - University College London (United Kingdom)
Nilufer Sari Aslam - University College London (United Kingdom)
Adam Sykulski - Imperial College London (United Kingdom)
Abstract: Capturing nonlinear effects while preserving interpretability remains a key challenge in regression modeling. SplitWise, a novel extension of stepwise regression, introduces adaptive transformations of numeric predictors into binary threshold-based features using shallow decision trees. These transformations are only applied when they improve model fit based on Akaike or Bayesian information criteria (AIC/BIC). This approach enhances the flexibility of linear models without sacrificing their transparency. Implemented as an R package, SplitWise is validated on synthetic and real-world datasets. Compared to traditional stepwise and penalized regression, it consistently produces more parsimonious, interpretable, and generalizable models.
