GNN Expressive Power Studied via Global Readout
A new arXiv preprint (2604.22870) investigates the expressive power of message-passing aggregate-combine-readout graph neural networks (ACR-GNNs), focusing on first-order (FO) properties. The authors show that sum aggregation and readout enable GNNs to capture FO properties inexpressible in the logic C2 on both directed and undirected graphs, extending results by Hauke and Wałęga (2026). They also identify two ways to restore characterisability with respect to C2: limiting local aggregation without restricting global readout, or running ACR-GNNs on graphs of bounded degree but unbounded size.
Key facts
- arXiv:2604.22870v1
- Studies expressive power of ACR-GNNs
- Focuses on first-order properties
- Sum aggregation and readout suffice for FO properties beyond C2
- Extends results by Hauke and Wałęga (2026)
- Two methods to restore C2 characterisability
- Limiting local aggregation without restricting global readout
- Running ACR-GNNs on bounded-degree graphs
Entities
Institutions
- arXiv