Exploring Ensemble Filters for Software Defect Prediction  
Author Taghi M. Khoshgoftaar

 

Co-Author(s) Kehan Gao; Lofton A. Bullard

 

Abstract Software quality prediction is a process of using software metrics gathered during system development process along with various data mining techniques to classify program models as either fault-prone (fp) or not-fault-prone (nfp). Two problems that often arise in software measurement datasets are high dimensionality (a dataset having too many independent attributes) and class imbalance (a dataset having many more members of one class than the other in a two-class classification problem). This paper presents several forms of the repetitive feature selection techniques to resolve these two issues. The general process of the repetitive feature selection method is that using a data sampling technique to alter the class ratio of the original dataset, then applying a feature ranking technique to the sampled dataset and ranking all the features according to their predictive capabilities. These two steps are repeated k times and results are combined using the mean aggregation. Various forms would be generated in this process when different filters (ranking techniques) and sampling methods are adopted. We study three sampling approaches (random undersampling (RUS), random oversampling (ROS) and synthetic minority oversampling (SMOTE)), each in conjunction with one of ten individual filters (rankers) and an ensemble filter that combine all of the filters. After feature election, there are two possible ways to form the training dataset – applying the selected features to the original data and applying the selected features to the sampled data. We investigate both and apply the repetitive feature selection process to three software datasets. We use Naïve Bayes to build classification models. The experimental results demonstrate that the ensemble filter makes better prediction than most individual filters and their average. Classification models built on the sampled datasets using the SMOTE technique show better prediction performance than those models built on the original datasets. However, this may not be the case when a different sampling technique like RUS or ROS is employed.

 

Keywords
   
    Article #:  19101
 
Proceedings of the 19th ISSAT International Conference on Reliability and Quality in Design
August 5-7, 2013 - Honolulu, Hawaii, U.S.A.