Integrated Topic Modeling and User Interaction Enhanced WebAPI Recommendation using Regularized Matrix Factorization for Mashup Application Development

Mashup application developers combine relevant web APIs from existing sources. Still, developers often face challenges in finding appropriate web APIs as they have to go through thousands of available ones. Recommending relevant web APIs might help, but very low API invocation from mashup applications creates a sparse dataset for the recommendation models to learn about the mashups and their invocation pattern, ultimately affecting their accuracy. Effectively reducing sparsity and using supplemental information such as mashup and web API specific features that trigger mashups to invoke the same web APIs in their applications and web APIs to be used together by a mashup can help to generate more accurate and useful recommendations. In this work, we developed a novel web API recommendation model for mashup application, which uses two-level topic modeling of mashups and user interaction with mashup and web APIs sequentially to reduce the sparsity of the initial dataset. Then, we applied regularized matrix factorization with the mashup and web API embeddings. These embeddings integrate 'mashup to mashup' and 'web API to web API' relationships with 'mashup to web API' invocation analysis. Compared with existing web API recommendation models, our model achieved 54% more precision, 36.4% more Normalized Discounted Cumulative Gain (NDCG), and 36% more recall value over other baseline models on a dataset collected from programmableWeb1.