KAN-GNN: Kolmogorov-Arnold Network Inspired Graph Neural Networks for Personalized API Recommendation

The rapid expansion of the Application Pro-gramming Interface (API) ecosystem has made the selection of appropriate APIs a critical challenge in mashup devel-opment. However, existing Graph Neural Network (GNN)-based approaches commonly rely on fixed and homogeneous API graphs, which limits their ability to provide personal-ized recommendations. To address this issue, we propose Kolmogorov-Arnold Network (KAN)-GNN, a novel frame-work inspired by Kolmogorov–Arnold Networks. We for-mulate API recommendation as a link prediction task on a mashup–API bipartite graph, explicitly incorporating person-alized requirements. At its core, KAN-GNN comprises two key components, both built upon a residual gated Rectified Linear Unit (ReLU)-KAN architecture: (1) a convolutional encoder that adaptively captures complex nonlinear inter-actions, and (2) a decoder that directly learns a flexible scoring function from node embeddings, thereby overcom-ing the rigidity and limited expressiveness of conventional encoder–decoder frameworks. Extensive experiments on the ProgrammableWeb dataset demonstrate that KAN-GNN con-sistently outperforms state-of-the-art baselines with notable gains across multiple evaluation metrics.