Abstract

Developing high-quality software products is directly related to the absence of unseen faults. The software fault prediction (SFP) methods have been considered to reduce the overall testing costs and discover probable faults. Various machine learning and deep learning algorithms have been used to predict faults. However, both deliver varying results regarding SFP. This research proposes a novel hybrid model (FR-CNN) based on a fine-tuned fully connected deep neural network, a random forest, and a convolutional neural network. Harnessing the strengths of combining these techniques, the proposed model aims to enhance prediction performance while reducing overfitting, resulting in a robust framework. We conduct extensive empirical studies to illustrate the effectiveness of the proposed model in predicting faults on six projects of the BugHunter dataset. The empirical results demonstrate that FR-CNN achieves competitive fault prediction performance while requiring lower computational complexity and training time than recent deep-learning approaches. In particular, it improves the best previously reported traditional machine-learning result on the BugHunter benchmark by up to 19.95% in F1-score.

Key words: Software fault prediction, Class imbalance, Deep learning, Ensemble learning, Machine learning, Hyperparameters