ABSTRACT With the explosive growth in the number of Web APIs, selecting an appropriate combination of APIs from massive services for mashup development has become an extremely challenging task. Existing research on API recommendation is mostly limited to a one‐shot, static recommendation mode, which struggles to accurately capture developers' vague requirement intentions and lacks an interactive mechanism to capture users' dynamic feedback. To this end, this paper proposes a novel interactive Web API recommendation framework named iLLMRec. First, the framework leverages a Large Language Model (LLM) to perform semantic enhancement on raw requirements, improving the clarity and completeness of descriptions. Subsequently, a feature extraction module is constructed by combining GloVe, CNN, and the self‐attention mechanism to deeply fuse the complex semantic correlations between requirements and APIs. To realize the dynamic optimization of recommendations, we innovatively introduce an interactive mechanism based on multi‐round reward feedback, and further transform API and category recommendations into a multi‐label classification problem. With the Binary Cross‐Entropy Loss (BCELoss) as the core loss function, the model parameters are adjusted in real time according to developers' selection behaviors. Experimental results on the real ProgrammableWeb dataset show that iLLMRec significantly outperforms existing state‐of‐the‐art methods in core metrics such as Precision, Recall, F1‐score, and NDCG. Experiments demonstrate that the combination of LLM‐based semantic enhancement and interactive feedback can effectively improve the accuracy of Web API recommendations, providing efficient decision support for complex mashup development.
Kang et al. (2026) studied this question.