SCALE-GNN: Advanced Graph Reduction for Scalable Test Case Prioritization Using GNN

Abstract

Test Case Prioritization (TCP) is crucial for efficient
regression testing, especially in CI pipelines where rapid feed
back is critical. While Graph Neural Network (GNN) methods
outperform traditional prioritization heuristics, their scalability is
severely limited on large software systems due to exploding graph
size and GPU memory requirements. This paper introduces
Scale-GNN, a scalable and fault-aware GNN framework that
aggressively reduces graph size through a three-stage pipeline
combining neural feature compression, fault-centric pruning, and
spectral graph coarsening.
Experiments across two datasets with 150 and 5000 test cases
demonstrate that Scale-GNN reduces graph size by up to 78%,
cuts training time by 3.1x, reduces GPU memory by 61%, and
preserves or improves fault detection effectiveness (APFD 0.85
0.93). Notably, Scale-GNN detects up to 88% of faults within
the first 10% of test executions, outperforming both classical
heuristics and unreduced GCN baselines. These results establish
Scale-GNN as a practical approach for integrating GNN-driven
prioritization into industrial-scale regression testing.