Shift-Invariant Attribute Scoring for Kolmogorov-Arnold Networks via Shapley Value

The article introduces ShapKAN, a pruning framework for Kolmogorov-Arnold Networks (KANs) that utilizes Shapley value attribution to assess node importance in a shift-invariant manner. This method addresses the limitations of traditional magnitude-based pruning techniques, which are sensitive to input coordinate shifts, thereby enhancing the interpretability and compression of KANs without sacrificing performance. The implications for practitioners include improved model efficiency and deployability in resource-constrained settings while maintaining a clear understanding of feature contributions.