Function-Guided Extended Latent Dirichlet Allocation Model for Complementary Cloud API Recommendation in Mashup Development

Abstract

In the cloud era, cloud application programming interface (API), as the best carrier for service delivery, capability replication, and data output, has become the core element of service-oriented software development. The existing cloud API recommendation methods adhere to a common paradigm: leveraging perceived quality of service and keyword matching to generate high-quality, single-function results, while overlooking the objective needs for function-guided complementary cloud APIs in service-oriented software development. Function-guided complementary cloud API recommendation aims to generate cloud APIs that are frequently co-invoked in conjunction with those API having given function, thereby satisfying the joint interests of developers. To this end, we proposed a function-guided extended latent Dirichlet allocation (ELDA) model for complementary cloud API recommendation. Specifically, we first conduct an analysis of real-world data from the cloud API ecosystems to illustrate both the necessity for complementary cloud API recommendations and the objective existence of a head effect within these APIs. Then we conceptualize the complementary relationship between a function and cloud APIs by treating the function as a document, with the corresponding cloud APIs represented as words within that document. Furthermore, we extend the classic latent Dirichlet allocation model by introducing two additional factors: (1) cloud API popularity and (2) functional sensitivity. These factors are designed to capture head effects within complementary cloud APIs. Additionally, we train both a positive and a negative ELDA model using the respective positive and negative corpus sets obtained. Furthermore, complementary cloud APIs relevant to the targeted function are generated by integrating the results from both the positive and negative ELDA models. Finally, experiments were conducted on two real-world cloud API datasets. The results demonstrate that the performance of ELDA surpasses that of the comparative methods. Sensitivity analysis of hyperparameters and case study findings further validate the effectiveness and practicality of ELDA.