Enhancing Developer Recommendations with Data Augmentation and Meta-Learning under Imbalanced Task Distributions

Accurate recommendation of appropriate developers for Pull Request (PR) reviews is crucial for en-suring efficient software maintenance in GitHub. However, existing studies primarily focus on developers’ review experience, overlooking the issue of data imbalance, which affects the effectiveness of the recommendations, i.e., imbalanced developer workload and reduced recommendation performance. To address this challenge, we pro-pose DAMRec, a developer recommendation method based on data augmentation and meta-learning frameworks. The key point of our approach is artificially diversifying the imbalanced training data, ensuring the model encounters a broader range of developers. In terms of model design, we provide local and global data augmentation strategies to increase the exposure of inactive developers and enhance their chances of being recommended. In the training process, we introduce a meta-learning based training framework to mitigate the impact of data imbalance on model training. Through extensive experiments on popular open-source GitHub projects, we demonstrate that DAMRec outperforms state-of-the-art methods in both recommendation accuracy and mean reciprocal rank (MRR). Com-pared to existing approaches, DAMRec provides more accurate developer recommendations and alleviates the problem of overburdening a small subset of developers with review tasks.