An Extended Software Defect Prediction Framework for Improving Accuracy

Abstract

Ensuring the quality and reliability of software systems is a critical challenge, mainly due to the difficulty in accurately predicting software defects. Traditional models often struggle with maintaining high accuracy across diverse datasets and classifiers. This study addresses this challenge by exploring the enhancement of defect prediction accuracy by integrating L2 regularisation with feature selection and dataset resampling techniques, forming the With Feature Selection and Regularization (WFS+RG) model.

The evaluation was conducted using five classifiers – Multilayer Perceptron, Bayes Net, Lazy IBK, J48, and Logistic Regression – applied to two datasets, PC1 and JM1, within the Waikato Environment for Knowledge Analysis (WEKA). The results indicate that the WFS+RG model consistently outperforms the traditional model With Feature Selection WFS only across most classifiers. Specifically, on the PC1 dataset, Lazy IBK achieved an accuracy of 97.29%, J48 reached 95.67%, and Multilayer Perceptron obtained 94.40%. Bayes Net and Logistic Regression also showed strong performances with 91.25 and 93.86% accuracy, respectively. On the JM1 dataset, Lazy IBK demonstrated significant improvement with an accuracy of 90.21%, while J48 and Multilayer Perceptron achieved 84.92 and 81.02%, respectively. Bayes Net and Logistic Regression reported 76.93 and 81.10% accuracy, respectively. However, the improvements are not uniform across all classifiers, with some showing minimal change. Statistical analysis conducted via Minitab confirms the robustness and significance of the WFS+RG model's impact, underscoring its potential as a more practical approach in software defect prediction tasks.