Generalization of the dependency-graph advantage with increasing system size

Determine whether the performance advantage provided by dependency-graph information in graph neural networks for code-to-architecture mapping generalizes as software system size increases beyond the single large-scale system evaluated.

Background

The study evaluates graph convolutional and graph attention networks against a graph-free multilayer perceptron on ten open-source systems. Although the graph-based models achieve a substantial advantage over the MLP on Chrome, Chrome is the only large-scale system in the benchmark, with 18,343 files. Consequently, the reported result cannot determine whether the observed benefit is a general scalability phenomenon or a system-specific effect. Establishing this would require evaluation across additional large software systems with reliable code-to-architecture ground truth mappings.

References

Because Chrome is the only large-scale system, this result alone cannot establish whether the graph advantage generalizes as system size increases.

— When Does the Dependency Graph Help Code-to-Architecture Mapping?  (2610.19769 - Jabeen et al., 14 Oct 2026) in Section 5, RQ1, subsection “How stable is mapping across runs?”