Is the Statistical Advantage Worth the Cost? An Empirical Comparison of KANs and MLPs for Structured Data Classification

This empirical benchmarking study compares Kolmogorov-Arnold Networks (KANs) and Multi-Layer Perceptrons (MLPs) on structured tabular classification tasks. Evaluating twelve public datasets across binary, multiclass, multilabel, and ordinal problems, the research utilizes standardized preprocessing and fixed hyperparameters. Performance is measured through test accuracy, F1-Score, paired hypothesis testing, and effect size analysis. The findings demonstrate that KANs statistically outperform MLPs in binary classification, examining if this statistical advantage justifies the cost of using KANs as an alternative architecture.