Context

The severity of a bug is often used as an indicator of how a bug negatively affects system functionality. It is used by developers to prioritize bugs which need to be fixed. The problem is that, for various reasons, bug submitters often enter the incorrect severity level, delaying the bug resolution process. Techniques that can automatically predict the severity of a bug can significantly reduce the bug triaging overhead. In our previous work, we showed that the accuracy of description-based severity prediction techniques could be significantly improved by using stack traces as a source of information.

Objective

In this study, we expand our previous work by exploring the effect of using categorical features, in addition to stack traces, to predict the severity of bugs. These categorical features include faulty product, faulty component, and operating system. We experimented with other features and observed that they do not improve the severity prediction accuracy. A Software system is composed of many products; each has a set of components. Components interact with each to provide the functionality of the product. The operating system field refers to the operating system on which the software was running on during the crash.

Method

The proposed approach uses a linear combination of stack trace and categorical features similarity to predict the severity. We adopted a cost sensitive K Nearest Neighbor approach to overcome the unbalance label distribution problem and improve the classifier accuracy.

Results

Our experiments on bug reports of Eclipse submitted between 2001 and 2015 and Gnome submitted between 1999 and 2015 show that the accuracy of our severity prediction approach can be improved from 5% to 20% by considering categorical features, in addition to stack traces.

Conclusion

The accuracy of predicting the severity of bugs is higher when combining stack traces and three categorical features, product, component, and operating system.

中文翻译：

---

使用堆栈跟踪和分类功能自动预测错误的严重性

语境

错误的严重程度通常用作指示错误如何负面影响系统功能的指标。开发人员使用它来确定需要修复的错误的优先级。问题在于，由于各种原因，错误提交者通常会输入错误的严重性级别，从而延迟了错误解决过程。可以自动预测错误严重性的技术可以大大减少错误分类的开销。在我们以前的工作中，我们表明通过使用堆栈跟踪作为信息源，可以大大提高基于描述的严重性预测技术的准确性。

目的

在这项研究中，我们通过探索除堆栈跟踪之外还使用分类功能来预测错误的严重性的作用，扩展了我们以前的工作。这些分类功能包括有缺陷的产品，有缺陷的组件和操作系统。我们对其他功能进行了实验，发现它们并未提高严重性预测的准确性。一个软件系统由许多产品组成。每个都有一组组件。组件相互交互以提供产品的功能。操作系统字段是指崩溃期间在其上运行软件的操作系统。

方法

所提出的方法使用堆栈跟踪和分类特征相似性的线性组合来预测严重性。我们采用了一种成本敏感的K最近邻方法来克服不平衡标签分配问题并提高分类器准确性。

结果

我们对2001年至2015年之间提交的Eclipse错误报告以及1999年至2015年之间提交的Gnome的错误报告进行的实验表明，除堆栈跟踪外，通过考虑分类特征，我们的严重性预测方法的准确性可以从5％提高到20％。

结论

将堆栈跟踪和三个分类功能（产品，组件和操作系统）结合使用时，预测错误严重性的准确性更高。