PRKG: Pre-Training Representation and Knowledge-Graph-Enhanced Web Service Recommendation for Mashup Creation

The number of online services is rapidly increasing due to the increased adoption of services-oriented technology. In this context, recommendation systems can provide high-quality Web services that meet Mashup developers’ expectations. The use of different kinds of auxiliary information in recommendation systems is commonplace. They enrich recommendation systems so that they can make relevant recommendations. Yet, knowledge graph-based service recommendation usually only considers the textual semantic information of the service and ignores the importance of discrete attribute information of the service for service recommendation. This may lead to the inability to comprehensively capture the multidimensional characteristics of services, thus affecting the accuracy and reliability of recommendations. To this end, this paper proposes a Web service recommendation method for Mashup creation that exploits pre-training representation and knowledge graphs as auxiliary information. Firstly, it uses the neural factorization machines and Doc2Vec to obtain the text semantic representation and the discrete attribute representation of Web services respectively. Secondly, it combines the text semantic representation and the discrete attribute representation to generate the pre-training representation as the input of knowledge graph convolutional networks. Thirdly, it constructs the Web services knowledge graph using Mashups, Web services, and related information and learns the preferences of Mashup developers and higher-order structural relations between Web services using knowledge graph convolutional networks to complete Web service recommendations. Finally, the proposed method is compared to the baselines, i.e., feature interaction-based (LR, FM, FFM, and NFM), KG-based (RippleNet and KGCN), and Doc2Vec for entity representation-based (DKGCN) Web service recommendation methods, using a real-world dataset from ProgrammableWeb. The experimental results show that the proposed method significantly improves the quality of recommendation in terms of the accuracy, recall, and Micro-F1.