Context:

Bug reports record issues found during software development and maintenance. A high-impact bug report (HBR) describes an issue that can cause severe damage once occurred after deployment. Identifying HBRs from the bug repository as early as possible is crucial for guaranteeing software quality.

Objective:

In recent years, many machine learning-based approaches have been proposed for HBR prediction, and most of them are based on supervised machine learning. However, the assumption of supervised machine learning is that it needs a large number of labeled data, which is often difficult to gather in practice.

Method:

In this paper, we propose hbrPredictor, which combines interactive machine learning and active learning to HBR prediction. On the one hand, it can dramatically reduce the number of bug reports required for prediction model training; on the other hand, it improves the diversity and generalization ability of training samples via uncertainty sampling.

Result:

We take security bug report (SBR) prediction as an example of HBR prediction and perform a large-scale experimental evaluation on datasets from different open-source projects. The results show: (1) hbrPredictor substantially outperforms the two baselines and obtains the maximum values of F1-score (0.7939) and AUC (0.8789); (2) with the dynamic stop criteria, hbrPredictor could reach its best performance with only 45% and 13% of the total bug reports for small-sized datasets and large-sized datasets, respectively.

Conclusion:

By reducing the number of required training samples, hbrPredictor could substantially save the data labeling effort without decreasing the effectiveness of the model.

中文翻译：

---

结合交互式机器学习和主动学习，提高高影响力的错误报告预测

内容：

错误报告记录了在软件开发和维护期间发现的问题。高影响力的错误报告（HBR）描述了一个问题，该问题一旦部署后就可能导致严重损坏。尽早从错误存储库中识别HBR对保证软件质量至关重要。

目的：

近年来，已经提出了许多基于机器学习的方法来进行HBR预测，并且大多数方法都是基于监督的机器学习。但是，监督式机器学习的假设是它需要大量的标记数据，在实践中通常很难收集这些数据。

方法：

在本文中，我们提出了hbrPredictor，它将交互式机器学习和主动学习结合到HBR预测中。一方面，它可以大大减少预测模型训练所需的错误报告的数量；另一方面，它通过不确定性采样提高了训练样本的多样性和泛化能力。

结果：

我们以安全错误报告（SBR）预测作为HBR预测的示例，并对来自不同开源项目的数据集进行大规模的实验评估。结果表明：（1）hbrPredictor明显优于两个基线，并获得F1分数（0.7939）和AUC（0.8789）的最大值；（2）使用动态停止条件，对于小型数据集和大型数据集，hbrPredictor可能仅在错误报告总数中分别只有45％和13％达到其最佳性能。

结论：

通过减少所需训练样本的数量，hbrPredictor可以在不降低模型有效性的情况下大大节省数据标记工作。