Abstract
Android developers frequently update source code to improve the performance, security, or maintainability of Android apps. Such Android code updating activities are intuitively repetitive, manual, and time-consuming. In this paper, we propose AutoUpdate, a Transformer-based automated code update recommendation approach for Android Apps, which takes advantage of code abstraction (Abs) and Byte-Pair Encoding (BPE) techniques to represent source code. Since this is the first work to automatically update code in Android apps, we collect a history of 209,346 updated method pairs from 3,195 real-world Android applications available on Google Play stores that span 14 years (2008-2022). Through an extensive experiment on our curated datasets, the results show that AutoUpdate(1) achieves a perfect prediction of 25% based on the realistic time-wise evaluation scenario, which outperforms the two baseline approaches; (2) gains benefits at least 17% of improvement by using both Abs and BPE; (3) is able to recommend code updates for various purposes (e.g., fixing bugs, adding new feature, refactoring methods). On the other hand, the models (4) could produce optimistically high accuracy due to the unrealistic evaluation scenario (i.e., random splits), suggesting that researchers should consider time-wise evaluation scenarios in the future; (5) are less accurate for a larger size of methods with a larger number of changed tokens, providing a research opportunity for future work. Our findings demonstrate the significant advancement of NMT-based code update recommendation approaches for Android apps.
Abstract (translated)
URL
https://arxiv.org/abs/2209.07048
