A Novel Approach for Converting N-Dimensional Dataset into Two Dimensions to Improve Accuracy in Software Defect Prediction

摘要

Software defect prediction model is trained using code metrics and historical defect information to identify probable software defects. The accuracy and performance of a prediction model largely depend on the training dataset. In order to provide proper training dataset, it is required to make the dataset clustered with less variabilities using clustering algorithms. However, clustering process is hampered due to multiple attributes of dataset such as Coupling between Objects, Response for Class, Lines of Code, etc. This research will aim to predict software defects through reducing code metrics dimensions to two latent variables. It will finally help the clustering algorithms to group data properly for the defect prediction model. In this paper, the dataset similarities are analyzed by reducing code metrics’ attributes into two latent variables based on their impacts to defects. Their impacts to defects can be analyzed using regression analysis because it identifies the relationship among a set of dependent and independent variables. Then, the code metrics are merged into two variables - PosImpactValue and NegImpactValue based on their positive or negative impact, respectively. As a result, multi-dimensional dataset is mapped into two-dimensional dataset. Plotting those dimensions reduced datasets enable distance-based clustering algorithms to group those datasets based on their similarities. Experiments have been performed on 18 releases of 6 open source software datasets such as jEdit, Ant, Xalan, Synapse, Tomcat and Camel. For comparative analysis, one of the most commonly used dimension reduction techniques named Principle Component Analysis (PCA) and two popular clustering techniques in defect prediction – DBSCAN and WHERE have been used in the experiment. First, the dimensions of the experimental datasets have been reduced using the proposed technique and PCA separately. Then, the reduced datasets have been clustered using DBSCAN and WHERE independently for identifying number of defects accurately. The comparative result analysis shows that the defect prediction models based on the clustering algorithms are more accurate for the dataset reduced by the proposed technique than PCA.