Reproduce “DynaTab: Dynamic Feature Ordering as Neural Rewiring for High-Dimensional Tabular Data” on OpenResearch

About this paper

Al Zadid Sultan Bin Habib, Gianfranco Doretto, Donald A. Adjeroh

Researchers from West Virginia University and The University of Utah developed DynaTab, an end-to-end deep learning model that employs dynamic feature ordering (DFO) inspired by neural rewiring to structure high-dimensional tabular data. This approach demonstrated strong performance on high-dimensional and complex mixed-regime tabular datasets, often ranking first or second and outperforming many baseline models in these challenging scenarios.

Key insights

• Drawing inspiration from neuroplasticity and neural rewiring, an adaptive feature ordering mechanism can dynamically adjust feature relevance and connectivity to optimize processing for specific tabular tasks.
• Learning an optimal, data-driven feature permutation can significantly reduce redundancy and reveal critical long-range dependencies, providing a strong structural inductive bias for sequential deep learning models.
• The utility of dynamic feature ordering can be predicted using metrics like the Intrinsic Dimensionality Factor, indicating that ordering is more beneficial for datasets with higher redundancy and complexity.

Summary by alphaXiv.