A Benchmarking Study of Kolmogorov-Arnold Networks on Tabular Data

TL;DR

This benchmarking study evaluates Kolmogorov-Arnold Networks (KANs) versus MLPs on real-world tabular datasets. It demonstrates that KANs provide competitive or superior accuracy and F1 scores, particularly on larger datasets, by leveraging edge-wise learnable activations and deeper representations inspired by Kolmogorov-Arnold theory. However, these gains come with higher computational costs, as reflected in increased FLOPS and longer training times. The findings suggest KANs are a promising alternative to MLPs for complex tabular data, with future work extending to regression tasks and broader data modalities.

Abstract

Kolmogorov-Arnold Networks (KANs) have very recently been introduced into the world of machine learning, quickly capturing the attention of the entire community. However, KANs have mostly been tested for approximating complex functions or processing synthetic data, while a test on real-world tabular datasets is currently lacking. In this paper, we present a benchmarking study comparing KANs and Multi-Layer Perceptrons (MLPs) on tabular datasets. The study evaluates task performance and training times. From the results obtained on the various datasets, KANs demonstrate superior or comparable accuracy and F1 scores, excelling particularly in datasets with numerous instances, suggesting robust handling of complex data. We also highlight that this performance improvement of KANs comes with a higher computational cost when compared to MLPs of comparable sizes.

Figures (2)

• Figure 1: Accuracy scores of KANs and MLPs as the number of parameters increases. For each dataset, we report the average accuracy across five runs for each parameter count.
• Figure 2: Number of FLOPS. The number of FLOPS performed by each model as the number of parameters increases is reported in megaflops (MFLOPS).