The past 10 years have seen the prediction of software defects proposed by many researchers using various metrics based on measurable aspects of source code entities (e.g. methods, classes, files or modules) and the social structure of software project in an effort to predict the software defects. However, these metrics could not predict very high accuracies in terms of sensitivity, specificity and accuracy. In this chapter, we propose the use of machine learning techniques to predict software defects. The effectiveness of all these techniques is demonstrated on ten datasets taken from literature. Based on an experiment, it is observed that PNN outperformed all other techniques in terms of accuracy and sensitivity in all the software defects datasets followed by CART and Group Method of data handling. We also performed feature selection by t-statistics based approach for selecting feature subsets across different folds for a given technique and followed by the feature subset selection. By taking the most important variables, we invoked the classifiers again and observed that PNN outperformed other classifiers in terms of sensitivity and accuracy. Moreover, the set of ‘if- then rules yielded by J48 and CART can be used as an expert system for prediction of software defects.
TopIntroduction
Machine learning techniques have been dominating in the last two decades. The recently published comprehensive state-of-the-art review (Mohanty et al., 2010) justifies this issue. The ability of software quality models to accurately identify critical faulty components allows for the application of focused verification activities ranging from manual inspection to automated formal analysis methods. Therefore, software quality models to ensure the reliability of the delivered products. Accurate prediction of fault prone modules enables the verification and validation activities that includes quality models: Musa, 1998, logistic regression (Basili et al., 1996), discriminant analysis (Khoshgoftaar, 1996), the discriminant power techniques (Schneidewind, 1992), artificial neural network (Khoshgoftaar, 1995), genetic algorithm (Azar et al., 2002), and classification trees (Gokhale et al., 1997; Khoshgoftar et al., 2002; Selby et al., 1988; Fenton et al., 1999).
A wide range of modeling techniques has been proposed and applied for software quality predictions. These include: proposed the Bayesian belief network as the most effective model to predict software quality.
Classification is a popular approach to predict software defects and involves categorizing modules, which is represented by a set of metrics or code attributes into fault prone (fp) non fault prone (nfp) by means of a classification model derived from data (Lessman et al., 2008), statistical methods (Basili et al., 1996; Khoshgoftar & Allen, 1999), tree based methods, (Guo et al., 2004; Khoshgoftar et al., 2000; Menzies et al., 2004; Porter et al., 1990; Selby et al., 1988), neural networks (Khoshgoftar et al., 1995, 1997) and analogy based approaches (El-Emam et al., 2001; Ganeshan et al., 2000; Khoshgoftar et al., 2003), Decision tree (Selby et al., 1988). The discriminative power techniques correctly classified 75 out of 81 fault free modules, and 21 out of 31 faulty modules (Porter et al., 1992). Lessmann et al., (2008) used 10 software development datasets from NASA MDP repository to predict software defects. Most recently, Pendharkar (2010) used the same dataset to test the efficacy of their hybrid exhaustive search and probabilistic neural network (PNN), and simulated annealing (SA) method.