This paper experimentally evaluates the performance of Kolmogorov Arnold Networks (KANs) on imbalanced data classification. The performance of KANs and Multi-Layer Perceptrons (MLPs) is compared and analyzed using 10 benchmark datasets. As a result, KANs, unlike MLPs, showed excellent performance on imbalanced data even without a resampling strategy, but existing imbalanced data processing strategies (resampling, focal loss) degraded the performance of KANs while having only a minor effect on the performance of MLPs. In addition, KANs had very high computational costs, but their performance improvement was limited. Statistical verification results showed that MLPs applying imbalanced data processing techniques achieved similar performance to KANs (|d| < 0.08) with less resource consumption. Therefore, KANs can be considered a specialized solution for raw imbalanced data when resources are sufficient, but their high resource consumption compared to their performance and low compatibility with existing resampling techniques limit their practical application. Future research directions include structural improvement of KANs for imbalanced learning, improved computational efficiency, and theoretical harmony with data augmentation.
