Micro-interaction Metrics Based Software Defect Prediction with Machine Learning, Immune Inspired and Evolutionary Classifiers: An Empirical Study

Abstract

Software developer’s pattern of activities, level of understanding of the source code and work practices are important factors that impact the defects introduced in software during development and its post-release quality. In very recent previous research (Lee et al. in Micro interaction metrics for defect prediction, pp 311–321, 2011), process metrics and micro-interaction metrics (Lee et al. in Micro interaction metrics for defect prediction, pp 311–321, 2011) that capture developer’s interaction with the source code have been shown to be influential on software defects introduced during development. Evaluation and selection of suitable classifiers in an unbiased manner is another conspicuous research issue in metrics based software defect prediction This study investigates software defect prediction models where micro-interactions metrics (Lee et al. in Micro interaction metrics for defect prediction, pp 311–321, 2011) are used as predictors for ten Machine Leaning (ML), fifteen Evolutionary Computation (EC) and eight Artificial Immune recognition system (AIRS) classifiers to predict defective files of three sub-projects of Java project Eclipse. They are -etc, mylyn and team. While no single best classifier could be obtained with respect to various accuracy measures on all datasets, we recommend a list of learning classifiers with respect to different goals of software defect prediction (SDP). For overall better quality of classification of defective and non-defective files, measured by F-measure, ensemble methods-Random Forests, Rotation Forests, a decision tree classifier J48 and UCS an evolutionary learning classifier system are recommended. For risk-averse and mission critical software projects defect prediction, we recommend logistic, J48, UCS and Immunos-1, an artificial immune recognition system classifier. For minimizing testing of non-defective files, we recommend Random Forests, Rotation Forests, MPLCS (Memetic Pittsburgh Learning Classifier) and Generational Genetic Algorithm (GGA) classifier.