Software Fault Prediction: A Systematic Mapping Study

Abstract

Context: Software fault prediction has been an important research topic in the software engineering field for more than 30 years. Software defect prediction models are commonly used to detect faulty software modules based on software metrics collected during the software development process. Objective: Data mining techniques and machine learning studies in the fault prediction software context are mapped and characterized. We investigated the metrics and techniques and their performance according to performance metrics studied. An analysis and synthesis of these studies is conducted. Method: A systematic mapping study has been conducted for identifying and aggregating evidence about software fault prediction. Results: About 70 studies published from January 2002 to December 2014 were identified. Top 40 studies were selected for analysis, based on the quality criteria results. The main metrics used were: Halstead, McCabe and LOC (67.14%), Halstead, McCabe and LOC + Object-Oriented (15.71%), others (17.14%). The main models were: Machine Learning(ML) (47.14%), ML + Statistical Analysis (31.42%), others (21.41%). The data sets used were: private access (35%) and public access (65%). The most frequent combination of metrics, models and techniques were: Halstead, McCabe and LOC + Random Forest, Naive Bayes, Logistic Regression and Decision Tree representing the (60%) of the analyzed studies. Conclusions: This article has identified and classified the performance of the metrics, techniques and their combinations. This will help researchers to select datasets, metrics and models based on experimental results, with the objective to generate learning schemes that allow a better prediction software failures.