TabKAN: Advancing Tabular Data Analysis using Kolmogorov-Arnold Network

TL;DR

TabKAN introduces a family of Kolmogorov–Arnol’d Network architectures tailored for tabular data, combining modular KAN variants with neural architecture search and a transfer-learning framework. Its core innovations are edge-wise learnable activations that enable intrinsic interpretability, a KAN-Mixer architecture for enhanced feature interactions, and a GRPO-based fine-tuning strategy for cross-domain knowledge transfer. Empirical results across ten public datasets show that TabKAN variants outperform classical methods and are competitive with, or surpass, Transformer-based models in both supervised and transfer learning tasks, with notable interpretability advantages over post hoc explanation methods. The work highlights TabKAN as a robust, efficient, and interpretable bridge between traditional ML and deep learning for structured data, with clear paths for future self-supervised and physics-informed extensions.

Abstract

Tabular data analysis presents unique challenges that arise from heterogeneous feature types, missing values, and complex feature interactions. While traditional machine learning methods like gradient boosting often outperform deep learning, recent advancements in neural architectures offer promising alternatives. In this study, we introduce TabKAN, a novel framework for tabular data modeling based on Kolmogorov-Arnold Networks (KANs). Unlike conventional deep learning models, KANs use learnable activation functions on edges, which improves both interpretability and training efficiency. TabKAN incorporates modular KAN-based architectures designed for tabular analysis and proposes a transfer learning framework for knowledge transfer across domains. Furthermore, we develop a model-specific interpretability approach that reduces reliance on post hoc explanations. Extensive experiments on public datasets show that TabKAN achieves superior performance in supervised learning and significantly outperforms classical and Transformer-based models in binary and multi-class classification. The results demonstrate the potential of KAN-based architectures to bridge the gap between traditional machine learning and deep learning for structured data.

Figures (11)

• Figure 1: The structure of the TabKAN framework for tabular datasets.
• Figure 2: Comparison of model interpretability between the built-in function-based explanations from TabKAN and a baseline using Partial Dependence Plots (PDP). TabKAN provides direct, parameterized feature-level insights, while PDP relies on post hoc approximations that may overlook complex interactions.
• Figure 3: Feature Importance based on ChebyKAN
• Figure 4: Feature Importance based on Fourier KAN
• Figure 5: AUC vs Percentage of Top Selected Features for Fourier KAN