An integrated graph neural network model for joint software defect prediction and code quality assessment.

Area of Science:

• Software Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Current software defect prediction and code quality assessment are often treated as independent tasks, missing opportunities for synergistic improvements.
• Existing graph-based methods struggle to jointly model software structure and quality characteristics, limiting their predictive power.
• The complex interplay between code structure, defect patterns, and quality indicators remains inadequately captured by traditional approaches.

Purpose of the Study:

• To propose a novel integrated model for simultaneous software defect prediction and code quality assessment.
• To leverage multi-level graph representations and graph neural networks (GNNs) for enhanced analysis.
• To exploit complementary information between defect prediction and quality assessment through a dual-branch GNN architecture.

Main Methods:

• Construction of multi-level graph representations integrating Abstract Syntax Trees (AST), Control Flow Graphs (CFG), and Data Flow Graphs (DFG).
• Development of a dual-branch, attention-based GNN architecture for shared representation learning and multi-task optimization.
• Simultaneous modeling of defect prediction and code quality assessment tasks to capture interdependencies.

Main Results:

• Achieved F1-scores of 0.811 and AUC values of 0.896 for defect prediction on real-world software projects.
• Demonstrated a 9.3% average improvement in code quality assessment accuracy across multiple dimensions.
• Significant performance gains over traditional independent methods, validating the integrated approach.

Conclusions:

• The proposed integrated model effectively captures complex structural dependencies and semantic relationships in source code.
• Jointly addressing defect prediction and quality assessment enhances analytical capabilities and provides actionable insights.
• Establishes a foundation for intelligent software engineering tools offering comprehensive code analysis.