Multi-Task Joint Training for Interactive Web API Recommendation via Exploring Mashup-API Interactions and Functional Description Documents

With the rapid growth of Web APIs, developers face increasing challenges in identifying suitable APIs for Mashup creation. The existing works for interactive Web API recommendation mainly rely on Mashup-API interactions and feature interactions are not deeply fused in the recommendation model. Moreover, the sparsity nature of Mashup-API interactions is negelcted in graph representation learning. To address these issues, we propose MIWAR, a multi-task joint training for interactive Web API recommendation method that exploits Mashup-API interactions and functional description documents. MIWAR enhances representation learning by combining a long-tail supervised contrastive graph encoder (LTS-LightGCL) for structural semantics and a BERT-based module for textual semantics. The aligned representations are fused and fed into a self-attentional neural factorization machine (SANFM) to model complex feature interactions. A multi-task joint training strategy jointly optimizes representation and recommendation tasks to improve effectiveness and generalization. Comprehensive experiments on a real-world dataset demonstrate that MIWAR consistently outperforms the state-of-the-art baselines across multiple evaluation metrics, achieving up to 5.72% improvement in Recall@10, 5.61% in NDCG@10, 9.08% in Precision@10, 8.97% in F1 Score, and 1.80% in Tail@10.