More Than Deep Learning: post-processing for API sequence recommendation

Chi Chen

doi:10.1007/s10664-021-10040-2

More Than Deep Learning: post-processing for API sequence recommendation

2021-10-29

4

OA

AI

DOI：10.1007/s10664-021-10040-2

Original

Original request for help

Save

Abstract

En 中文

In the daily development process, developers often need assistance in finding a sequence of APIs to accomplish their development tasks. Existing deep learning models, which have recently been developed for recommending one single API, can be adapted by using encoder-decoder models together with beam search to generate API sequence recommendations. However, the generated API sequence recommendations heavily rely on the probabilities of API suggestions at each decoding step, which do not take into account other domain-specific factors (e.g., whether an API suggestion satisfies the program syntax and how diverse the API sequence recommendations are). Moreover, it is difficult for developers to find similar API sequence recommendations, distinguish different API sequence recommendations, and make a selection when the API sequence recommendations are ordered by probabilities. Thus, what we need is more than deep learning. In this paper, we propose an approach, named Cook, to combine deep learning models with post-processing strategies for API sequence recommendation. Specifically, we enhance beam search with code-specific heuristics to improve the quality of API sequence recommendations. We develop a clustering algorithm to cluster API sequence recommendations so as to make it easier for developers to find similar API sequence recommendations and distinguish different API sequence recommendations. We also propose a method to generate a summary for each cluster to help developers understand the API sequence recommendations. Our evaluation results have shown that (1) three deep learning models with our heuristic-enhanced beam search achieved better performance than with the original beam search in terms of CIDEr-1, CIDEr-5 and CIDEr-10 scores, with an average improvement of 1.8, 2.3 and 2.3, respectively; and (2) our clustering algorithm achieved high performance on six metrics and outperformed two variant clustering algorithms. Moreover, our user study with 24 participants shows that Cook can help developers accomplish programming tasks faster and pass more test cases, and the participants confirm that clusters and summaries indeed help them understand and select the correct API sequence recommendations.

Keywords:

API

Recommendation

Deep learning

Encoder-decoder

Post-processing

More Than Deep Learning: post-processing for API sequence recommendation

Abstract

Journal

Organization

Cited Papers