A Classification Framework for Software Defect Prediction Using Multi-filter Feature Selection Technique and MLP

Abstract

Production of high quality software at lower cost can be possible by detecting defect prone software modules before the testing process. With this approach, less time and resources are required to produce a high quality software as only those modules are thoroughly tested which are predicted as defective. This paper presents a classification framework which uses Multi-Filter feature selection technique and Multi-Layer Perceptron (MLP) to predict defect prone software modules. The proposed framework works in two dimensions: 1) with oversampling technique, 2) without oversampling technique. Oversampling is introduced in the framework to analyze the effect of class imbalance issue on the performance of classification techniques. The framework is implemented by using twelve cleaned NASA MDP datasets and performance is evaluated by using: F-measure, Accuracy, MCC and ROC. According to results the proposed framework with class balancing technique performed well in all of the used datasets.

Software Defect Prediction Using Variant based Ensemble Learning and Feature Selection Techniques

“…These datasets are currently available at [24]. In this research we have used the DS'' version of NASA datasets which is already been used by many researchers [1,2,3,4,5,25,26,27]. Second stage of the framework deals with the selection of best variants from different classifiers (Fig.3).…”

“…In Data preprocessing, two tasks are performed: Resampling [31,32] and Randomization. Resampling is performed to resolve the issue of class imbalance in the datasets as this issue can compromise the accuracy of proposed classification framework [2,3,4,5]. To perform this task, the builtin function of WEKA is used (weka.filters.supervised.instance.Resample).…”

“…Feature Selection activity is applied on all the datasets in order to select only those features which highly participate in the classification process. As, it has been proved now from the studies [2], [5] that those features should be removed from the dataset which do not participate in classification process as these features may reduce the performance. In this research, feature selection is performed by "Cfs Subset Evaluator" [28,29,30] with BestFirst search method, whereas full dataset is given for training.…”

“…Then a detailed comparison is performed with various widely used supervised classification techniques from research [1]. Researchers in [5] proposed a classification framework to predict the defective modules before the testing. They used a multi filter feature section technique with an aggregated method and Artificial Neural Network (MLP).…”

“…The prediction of software defects is a binary classification problem as we have to identify that the particular module is defective or non-defective. Many researchers have focused on machine learning techniques to solve the binary classification problems such as: Network Intrusion Detection [21,22], Sentiment Analysis [13,14,15,16,17,18], Rainfall Prediction [19,20], and Software Defect Prediction [1,2,3,4,5]. The process of software defect prediction has been focused by many researchers in the last decade however improving the prediction accuracy has always been the main concern.…”