KANs Outperform MLPs on Tabular Data, But at Higher Cost

Kolmogorov-Arnold Networks (KANs) have garnered significant interest as a potential alternative to traditional Multi-Layer Perceptrons (MLPs) due to their theoretical advantages in function approximation. This research provides an empirical comparison of KANs and MLPs specifically for structured tabular classification tasks, evaluating their out-of-the-box performance across twelve diverse public datasets. The study standardized preprocessing, architecture, and hyperparameter settings for both model types, assessing performance using test accuracy, F1-Score, and statistical hypothesis testing. The results indicate that KANs statistically outperform MLPs in binary and multiclass classification problems, demonstrating an aggregate advantage across all datasets. Despite this performance edge, the observed medium effect size (d = -0.46) prompts a crucial cost-benefit analysis. KANs achieve superior generalization through their adaptive spline-based mappings but incur significantly higher parameter counts and computational demands compared to MLPs. This suggests that while KANs are preferable for applications demanding high precision, MLPs remain a robust and efficient choice when computational resources are constrained.