Incremental learning of code authors over time

doi:10.1016/j.jss.2025.112527

Incremental learning of code authors over time

2025-06-17

0

PRE

AI

DOI：10.1016/j.jss.2025.112527

Original

Original request for help

Save

Abstract

En 中文

Identifying code authors is essential in many research topics, and various approaches have been proposed. Recent studies show that the temporal effect can significantly affect existing approaches: their trained models rapidly become outdated and ineffective due to the evolution of code styles over time. To our knowledge, only a recent approach tries to alleviate the temporal effect. This approach treats the temporal effect problem as a cross-domain problem and uses domain adaptation to reduce the temporal effect. Although this approach achieves promising results on their datasets, the evaluation of this approach shows that the effectiveness of transferred models decreases with the increasing intervals between the source and the target domains. In addition, an author can have much more or fewer files in real development. A recent study shows that the effectiveness of existing approaches is significantly reduced if they are evaluated on imbalanced data.

Incremental learning of code authors over time

Abstract

Journal

Organization