Two-Stage Temporal Graph Neural Network for Web API Recommendation in Mashup Creation

Mashup applications integrate multiple web APIs to deliver composite functionality that a single service cannot provide. As public API repositories grow to tens of thousands of services, choosing which APIs to combine for a new mashup has become a difficult recommendation problem because the search space is large, usage data are sparse, and the ecosystem evolves continuously over time. We address this problem with a two-stage recommendation framework built around a Temporal Graph Neural Network (TGN). First, a content-based retrieval stage uses pre-trained sentence embeddings of mashup and API descriptions, together with a time-aware filter, to select a small set of semantically relevant candidate APIs for each mashup. Second, a TGN model treats historical mashup–API interactions as a time-ordered graph of events and maintains a dynamic memory for every mashup and API. By updating these memories when new interactions occur, the model learns time-sensitive representations that reflect both recent usage and long-term relational patterns, and then ranks the candidate APIs using a pairwise ranking loss. Experiments on a real-world ProgrammableWeb dataset show that the combined two-stage model significantly improves Recall@K compared with both content-only and TGN-only variants, as well as with strong collaborative filtering and static graph baselines. The proposed architecture is modular and can be extended with richer metadata and additional node types, making it a promising foundation for future intelligent mashup development and other API recommendation scenarios.