Graph neural network-based mutation-aware regression test ordering using code dependency graphs and execution traces
The mutation-aware test prioritisation system in this paper uses Graph Neural Networks (GNNs) to combine static program structure, dynamic execution traces, and mutation coverage into a hybrid graph representation to enhance regression testing. The framework embeds higher-order dependencies in test cases using GCN, GAT, and GraphSAGE variations and ranks them using a multi-objective optimisation function that balances fault detection, execution cost, and mutation coverage. On benchmark datasets like Defects4J and ManySStuBs4J, the proposed approach consistently outperforms traditional baselines (coverage-based APFD = 72.4 %, cost-based = 74.5 %) and ML baselines (LSTM = 80.1 %, RL = 82.7 %), achieving an average APFD of 88.9 % and mutation score of 84.6 % with a 16.1-second execution overhead. Statistical tests (Wilcoxon signed-rank, p < 0.05) indicate the robustness of these gains. Ablation experiments show that removing execution traces or mutation characteristics reduces APFD by 5–8 %, emphasising their relevance. Qualitative research shows that GNN embeddings cluster fault-related test cases for interpretable prioritisation. The suggested paradigm for contemporary regression testing is scalable, accurate, and mutation-driven.• Multi-Tiered Graph-Based Architecture: The method transforms raw program artifacts (codebase, mutants, test traces) into Program Dependence Graphs and Call Graphs, where nodes represent program elements and edges capture dependencies enriched with runtime characteristics.• GNN-Powered Multi-Objective Optimization: Core innovation uses Graph Neural Networks (GCN, GAT, GraphSAGE) to create enriched embeddings through iterative neighborhood aggregation, feeding into a scoring function that balances fault detection potential, execution cost, and mutation coverage.• Superior Validated Performance: Achieves 88.9 % APFD compared to 82.7 % for best baseline methods on real-world datasets, with statistical significance confirmed through Wilcoxon signed-rank tests across multiple evaluation metrics.
Paper
Full text
Graph neural network-based mutation-aware regression test ordering using code dependency graphs and execution traces
Semantic Scholar · Computer Science · 2025
Abstract
The mutation-aware test prioritisation system in this paper uses Graph Neural Networks (GNNs) to combine static program structure, dynamic execution traces, and mutation coverage into a hybrid graph representation to enhance regression testing. The framework embeds higher-order dependencies in test cases using GCN, GAT, and GraphSAGE variations and ranks them using a multi-objective optimisation function that balances fault detection, execution cost, and mutation coverage. On benchmark datasets like Defects4J and ManySStuBs4J, the proposed approach consistently outperforms traditional baselines (coverage-based APFD = 72.4 %, cost-based = 74.5 %) and ML baselines (LSTM = 80.1 %, RL = 82.7 %), achieving an average APFD of 88.9 % and mutation score of 84.6 % with a 16.1-second execution overhead. Statistical tests (Wilcoxon signed-rank, p < 0.05) indicate the robustness of these gains. Ablation experiments show that removing execution traces or mutation characteristics reduces APFD by 5–8 %, emphasising their relevance. Qualitative research shows that GNN embeddings cluster fault-related test cases for interpretable prioritisation. The suggested paradigm for contemporary regression testing is scalable, accurate, and mutation-driven.
- Multi-Tiered Graph-Based Architecture: The method transforms raw program artifacts (codebase, mutants, test traces) into Program Dependence Graphs and Call Graphs, where nodes represent program elements and edges capture dependencies enriched with runtime characteristics.
- GNN-Powered Multi-Objective Optimization: Core innovation uses Graph Neural Networks (GCN, GAT, GraphSAGE) to create enriched embeddings through iterative neighborhood aggregation, feeding into a scoring function that balances fault detection potential, execution cost, and mutation coverage.
- Superior Validated Performance: Achieves 88.9 % APFD compared to 82.7 % for best baseline methods on real-world datasets, with statistical significance confirmed through Wilcoxon signed-rank tests across multiple evaluation metrics.
References (18)
Scroll for more · 6 remaining