API-Prefer: An API Package Recommender System Based on Composition Feature Learning

Yancen Liu, Jian Cao

Abstract: With the exponential increase in Web Application Programming Interfaces (APIs), selecting appropriate APIs to construct a mashup is a challenging task. When multiple APIs are put together, their overall function is not just a superposition of their individual functions in many cases. Unfortunately, the approaches proposed to date do not sufficiently model the synthetical functions of the combined APIs. In this paper, an API Package recommender system based on composition feature learning (API-Prefer) is proposed. API-Prefer tries to learn the composition features of an API pair. Then the composition features can be used to predict whether this API pair can be adopted by a mashup or not. Specifically, a deep neural network is designed for composition feature learning and adoption probability prediction in API-Prefer. Since there is a large amount of API pairs, API-Prefer applies a strategy to select the potential APIs first, then the API packages can be discovered based on the predicted scores over multiple API pairs. Experiments on a real-world dataset show API-Prefer is significantly better than the comparative methods.

External IDs:dblp:conf/icsoc/LiuC20