Deep Learning-Based Automated Software Defect Prediction Model
Jiarui Fan, Hong Chen · 2024
This paper introduces a novel software defect prediction model employing deep learning techniques to enhance the automation and accuracy of identifying software faults. Traditional defect prediction approaches often fall short in effectively parsing and understanding the intricate structures and semantics encoded in source codes. This study leverages Abstract Syntax Tree (AST) features, which excel in capturing the structural and semantic aspects of code. We meticulously detail the methodology for extracting and transforming these AST features into vectors suitable for analysis. A dedicated Convolutional Neural Network (CNN) architecture is then employed to interpret these features, focusing on model optimization through advanced parameter tuning and regularization techniques. The model's efficacy was rigorously tested using the Promise public dataset and a Google code base, demonstrating significant improvements over conventional methods. The results indicate enhancements in both the Area Under the Curve (AUC) and the F1-measure, achieving scores of 0.92 and 0.85, respectively. These findings underscore the potential of deep learning in revolutionizing automated software defect prediction, highlighting its superior ability to interpret and analyze complex code structures for more reliable defect identification.