Hybrid Techniques and Evidential Neural Networks for Optimizing Parameter Settings in Software Defect Prediction

Abstract

This paper provides a comprehensive exploration of hybrid approaches and empirical 
neural network (ENN) methodologies for parameter optimization in Software 
Defect Prediction (SDP) models. Traditional SDP models often struggle with 
challenges such as correlated metrics and inconsistent predictive performance due 
to suboptimal parameter settings. To address these issues, this study combines 
search-based optimization techniques with ENNs, demonstrating their efficacy in 
reducing metric correlations, enhancing prediction accuracy, and improving overall 
model robustness. The novel integration of Gaussian-randomized fuzzy numbers is 
highlighted as a significant advancement, allowing for dynamic adjustment to varying 
dataset characteristics and enhanced handling of uncertainty in model predictions. 
Extensive experimentation illustrates the ability of these hybrid approaches to 
manage complex, high-dimensional data environments while maintaining 
computational efficiency and scalability. This research also emphasizes the 
interpretability of predictive outcomes achieved through ENNs, making them 
suitable for real-world software engineering applications where explainability is 
critical. The findings underscore the transformative potential of adopting advanced 
hybrid models in SDP to achieve more accurate, reliable, and efficient defect 
prediction. This study delves into the application of hybrid approaches and empirical 
neural networks (ENNs) for optimizing parameter settings in Software Defect 
Prediction (SDP) models. By leveraging search-based optimization techniques 
combined with ENNs, this research demonstrates a significant reduction in 
correlated metric issues and a notable improvement in predictive accuracy. The 
integration of Gaussian-randomized fuzzy numbers emerges as a groundbreaking 
development, enabling robust model design and enhancing the interpretability of 
predictive outcomes. Detailed experimentation highlights the ability of these hybrid 
techniques to adapt dynamically to varying dataset complexities while maintaining 
high accuracy levels. These findings underscore the potential of advanced hybrid 
models for scalable and efficient defect prediction in software engineering.