Mechanism underlying the interaction between edge-girth structure and message passing

Identify why the interaction between per-edge cycle structure encoded by edge-girth and multiplicity and gated message passing produces the observed performance gain on graph regression beyond the contributions of the gated architecture and the descriptor considered separately.

Background

The experiments show that EGAGNN, which combines gated message passing with per-edge edge-girth and shortest-cycle multiplicity features, substantially outperforms the same architecture without structural descriptors on Zinc-12k. The authors’ ablation indicates that the gated architecture alone achieves a test MAE of approximately 0.204, while adding the descriptor accounts for the remaining improvement.

Although the paper bounds the separate contributions of the architecture and descriptor, it does not identify the mechanism responsible for their interaction. The unresolved issue concerns why combining per-edge cycle structure with message passing is so effective for the observed regression task.

References

What remains is a mechanism we can bound but not identify: the gated architecture alone reaches $0.204$ (Table~\ref{tab:descriptors}), the descriptor contributes the rest, and why the interaction of per-edge cycle structure with message passing is worth that much is open.

— Edge-Girth as a Structural Edge Feature for Graph Neural Networks  (2609.01441 - Marey et al., 1 Sep 2026) in Section 5, Discussion and Limitations, paragraph “How far the supervised evidence reaches.”