Cooperative Mashup Embedding Leveraging Knowledge Graph for Web API Recommendation

Abstract

Creating top-notch Mashup applications is becoming increasingly difficult with an overwhelming number of Web APIs. Researchers have developed various API recommendation techniques to help developers quickly locate the right API. In particular, deep learning-based solutions have attracted much attention due to their excellent representation learning capabilities. However, existing methods mainly use textual or graphical information, and do not fully consider the two, which may lead to suboptimal representation and damage recommendation performance. In this paper, we propose a Cooperative Mashup Embedding (CME) neural framework that integrates knowledge graph embedding and text encoding, using Node2Vec to convert entities into numerical vectors and BERT to encode text descriptions. A cooperative embedding method was developed to optimize the entire model while capturing graph and text data knowledge. In addition, the representations obtained by the framework of the three recommendation models are derived. Experimental results on the ProgrammableWeb dataset indicate that our proposed method outperforms the SOTA methods in recommendation performance metrics Top@{1,5,10}. Precision and Recall have increased from 3% to 11%, while NDCG and MAP have improved from 3% to 6%.