A novel deep neural network structure for software fault prediction.

Area of Science:

• Software Engineering
• Machine Learning
• Artificial Intelligence

Background:

• Software fault prediction is vital for identifying potential defects early in the development lifecycle.
• Existing machine learning and deep learning models face challenges like low accuracy, data imbalance, and overfitting.
• Large datasets are crucial for deep learning's superior performance, yet common datasets like NASA MDP are limited.

Purpose of the Study:

• To address the limitations of current fault prediction models.
• To explore the application of deep learning on the larger BugHunter dataset.
• To propose a novel deep neural network (DNN) structure for improved fault prediction.

Main Methods:

• A novel deep neural network (DNN) architecture utilizing convolutional layers was developed.
• The model was designed to handle class imbalance and overfitting issues inherent in fault prediction datasets.
• Extensive empirical evaluations were conducted, comparing the DNN against traditional machine learning, ensemble learning, and state-of-the-art deep learning models.

Main Results:

• The proposed DNN structure demonstrated significant improvements in predicting fault-prone methods across 15 BugHunter projects.
• The average F1-score was enhanced by 20.01%, indicating superior predictive performance.
• The model effectively addressed class imbalance and overfitting challenges.

Conclusions:

• Deep neural networks (DNNs) offer a practical and effective approach for software fault prediction.
• The developed DNN model shows significant potential for improving software reliability and optimizing development resources.
• Further research can leverage these findings for more robust fault prediction systems.