GCL-GCN: a Graph Contrastive LearningBased Graph Convolutional Network for API Recommendation

With the development of Service-Oriented Architecture (SOA), it has become a trendy to develop applications by combining APIs into Mashups. As the number of API grows rapidly, some API recommendation approaches have emerged in recent years and achieved some success. However, as one mashup only contains two or three APIs on average, the lack of interactive information results in poor accuracy in many cases. In view of this problem, we propose a graph contrastive learning based graph convolutional network approach for API recommendation, named GCL-GCN. Concretely, Singular Value Decomposition (SVD) is first used to reconstruct Mashup-API interaction graphs. Then Graph Convolutional Network (GCN) is applied to generate feature vectors of each Mashup and API nodes in the reconstructed and original interaction graph respectively. Thirdly, contrastive learning is implemented between the feature vectors to get loss function which is utilized to optimize the parameters of GCN. At last, API recommendation is generated according to the inner product of embedding vectors output from GCN. We conducted some experiments on an actual dataset, and the experimental results show that our approach outperforms some baseline methods, mitigating the problems caused by inaccurate contrast signals while preserving the global collaboration of the graphs.