A Systematic Review of Unsupervised Defect Prediction Dataset

This dataset is about a systematic review of unsupervised learning techniques for software defect prediction (our related paper: "A Systematic Review of Unsupervised Learning Techniques for Software Defect Prediction" in Information and Software Technology [accepted in Feb, 2020] ). We conducted this systematic literature review that identified 49 studies which satisfied our inclusion criteria containing 2456 individual experimental results. In order to compare prediction performance across these studies in a consistent way, we recomputed the confusion matrices and employed MCC as our main performance measure. From each paper we extracted: Title, Year, Journal/conference, 'Predatory' publisher? (Y | N), Count of results reported in paper, Count of inconsistent results reported in paper, Parameter tuning in SDP? (Yes | Default | ?) and SDP references(SDPRefs OrigResults | SDPRefs |SDPNoRefs | OnlyUnSDP). Then from within each paper, we extracted for each experimental result including: Prediction method name (e.g., DTJ48), Project name trained on (e.g., PC4), Project name tested on (e.g., PC4), Prediction type (within-project | cross-project), No. of input metrics (count | NA), Dataset family (e.g., NASA), Dateset fault rate (%), Was cross validation used? (Y | N | ?), Was error checking possible? (Y | N), Inconsistent results? (Y | N | ?), Error reason description (text), Learning type (Supervised | Unsupervised), Clustering method? (Y | N | NA), Machine learning family (e.g., Un-NN), Machine learning technique (e.g., KM), Prediction results (including TP, TN, FP, FN, etc.).本数据集围绕软件缺陷预测（Software Defect Prediction, SDP）的无监督学习技术系统性综述展开，相关研究论文为发表于《Information and Software Technology》、2020年2月录用的"A Systematic Review of Unsupervised Learning Techniques for Software Defect Prediction"。本次系统性文献综述共筛选出49项符合纳入标准的研究，涵盖2456项独立实验结果。为以统一标准对比各研究的预测性能，我们重新计算了混淆矩阵，并以马氏相关系数（Matthews Correlation Coefficient, MCC）作为核心性能评估指标。我们从每篇入选文献中提取了以下信息：文献标题、发表年份、刊载期刊/会议、是否为掠夺性出版商（Y|N）、文献报告的实验结果总数、文献报告的不一致结果数量、SDP参数调优情况（是|默认|？）以及SDP参考文献类型（SDPRefs OrigResults | SDPRefs | SDPNoRefs | OnlyUnSDP）。针对每篇文献中的每项实验结果，我们进一步提取了以下内容：预测方法名称（如DTJ48）、训练所用项目名称（如PC4）、测试所用项目名称（如PC4）、预测类型（项目内|跨项目）、输入指标数量（计数|NA）、数据集族（如NASA）、数据集缺陷率（%）、是否使用交叉验证（Y|N|？）、是否可开展误差检查（Y|N）、是否存在不一致结果（Y|N|？）、误差原因描述（文本）、学习类型（监督|无监督）、是否为聚类方法（Y|N|NA）、机器学习家族（如Un-NN）、机器学习技术（如KM）以及预测结果（包含TP、TN、FP、FN等指标）。