Shift-Invariant Attribute Scoring for Kolmogorov-Arnold Networks via Shapley Value
A new framework called ShapKAN has been developed to address the challenges of pruning Kolmogorov-Arnold Networks (KANs). This method utilizes Shapley values to evaluate node importance in a manner that is invariant to input coordinate shifts. By quantifying each node's true contribution, ShapKAN provides more reliable importance rankings than traditional magnitude-based pruning techniques. Experiments show that ShapKAN effectively compresses KANs while preserving their interpretability and performance, making them more suitable for resource-constrained applications. AI
IMPACT Enhances the interpretability and efficiency of KANs, potentially enabling wider adoption in resource-limited settings.