Abstract: Software defect prediction can effectively improve the efficiency of software test. Based on Na?ve Bayes statistical theory, a software defect prediction model named LGD-NB is put forward by using geometrical relationship between a point and a line. The LGD-NB model has two forms; the one is decision-making based on the least risk which has more accurate description to the cost of misclassification. By defining the high risk area of decision-making, the second form of LGD-NB is to be a meta-classifier which provides an ensemble scheme for software defect prediction by which the another classifier can classify the modules which locate at the high risk area. The experimental results show that the LGD-NB based on the least risk has better prediction performance than Na?ve Bayes models. The LGD-NB ensemble with SVM has also effectively improved the prediction performance.

Key words: software defect prediction, Na?Ve Bayes, line geometry division, meta-classifier, ensemble scheme

摘要： 缺陷预测能够有效地提升软件测试的效率。基于朴素贝叶斯理论，提出了一个利用平面中点与直线几何关系进行分类的软件缺陷预测模型LGD-NB。LGD-NB有两种工作模式，当其基于最小风险进行决策时，比传统的朴素贝叶斯具有对代价更为精确的描述；在定义了几何上的高风险决策区域后，LGD-NB可作为元分类器，提供一个可集成其他分类模型进行二次分类的集成框架。实验结果显示：基于最小风险LGD-NB模型的预测性能优于传统的朴素贝叶斯；而集成了SVM算法后的LGD-NB，其预测能力也有较为明显的提升。

关键词: 软件缺陷预测, 朴素贝叶斯, 直线集合分割, 元分类器, 集成框架