SOFTWARE DEFECT PREDICTION USING MACHINE LEARNING TECHNIQUES

Abstract

Software defect prediction is a process of classification which determines whether a software module is defective or not. A defect prediction model is a method where a set of independent variables (the predictors) are used to predict the value of a dependent variable (the defect-proneness of a class) using a machine learning classifier. Innumerable studies are present in literature that studies the effect of dimensionality reduction on performance of models developed for Software Defect Prediction. It is said to improve certain models. Also, Software defect prediction is a costly activity and the problem relies in the fact that many feature-extraction methods based on traditional as well as novel like deep learning are there for dimensionality reduction. So, it becomes very difficult to choose any method based on its working, its pros and cons and its performance for dimensionality reduction. Thus, there arises a need of comparison study for feature extraction technique which exists earlier in literature. This study aims to provide literature review on the previously existing feature reduction techniques in software defect prediction. The study helps software developers in identifying the commonly prevalent as well as novel feature extraction techniques, their characteristics and their performance in area of software defect prediction and guides the researchers in conducting future research. The comparison is performed on nine open-source software-systems written in Java using four mostly used feature extraction technique and a machine learning classifier. The model validation is performed by 10 fold cross validation method and the performance measure used is accuracy and ROCAUC. Results of the study indicate that autoencoders is an effective method to reduce the dimensions of a dataset successfully.