<p>Selecting suitable Web APIs for Mashup creation is increasingly challenging due to the vast number of available services, data sparsity, and low-quality descriptions. Existing methods often struggle to effectively handle the recommendation of low-quality descriptive data and the required functional similarity and complementarity. First, in the introduction part, this paper establishes the necessity of the research by clarifying the specific challenges in Web API recommendation. Next, the method part provides a targeted solution, where this paper proposes a novel framework combining Large Language Models (LLMs) and heterogeneous graph learning. The framework aims to deeply enhance Mashup and API descriptions using LLMs, while simultaneously constructing a Mashup-API heterogeneous information network by integrating the call relationships between Mashups and APIs and the associations between APIs and labels. It also extracts features using Bert encoding and Graph Convolutional Network (GCN) techniques, and integrates the extracted features into Neural and Attentional Factorization Machine (NAFM) to generate Top-K recommendations. Subsequently, the results part shows through experiments on the ProgrammableWeb dataset that our method significantly outperforms several state-of-the-art methods, and ablation studies confirm the key contributions of the method. Finally, the discussion part elaborates on its theoretical contributions and practical significance. Our main innovations include: utilizing LLMs to enhance service descriptions; constructing a heterogeneous graph capable of capturing similarity and complementarity relationships; and developing a comprehensive recommendation framework that combines semantic and structural information.</p>