Revisiting Chebyshev Polynomial and Anisotropic RBF Models for Tabular Regression

The article presents the development of new smooth-basis models for tabular regression, including an anisotropic RBF network with data-driven center placement, a ridge-regularized Chebyshev polynomial regressor, and a smooth-tree hybrid model. These models, released as scikit-learn-compatible packages, were benchmarked against tree ensembles and a pre-trained transformer across 55 datasets, revealing that while the transformer achieved the highest accuracy, the smooth models demonstrated tighter generalization gaps in CPU-based environments. This suggests that incorporating smooth-basis models could enhance performance in applications where gradual response variation is beneficial.