Core Concepts
The authors investigate the expressivity of different versions of graph neural networks using modal and guarded fragments of first-order logic with counting. They aim to determine if 2-GNNs are more powerful than 1-GNNs.
Abstract
The content delves into the architecture and operation of Graph Neural Networks (GNNs), focusing on message passing algorithms on graph vertices. It compares two versions of GNNs: 1-sided and 2-sided, discussing their expressivity in terms of first-order logic with counting. The study explores whether targeted messages in 2-GNNs offer superior computational capabilities compared to 1-GNNs.
The authors establish that both versions have been used practically, but theoretical work has predominantly focused on 1-GNNs. They address the question of whether the two versions differ in their ability to compute functions, emphasizing non-uniform expressivity. By proving results related to uniform and non-uniform settings, they highlight the varying computational power between 1-GNNs and 2-GNNs.
Furthermore, the content introduces modal and guarded fragments in logic that correspond to message passing mechanisms in GNNs. It discusses how these fragments impact the expressive power of GNN models over labelled undirected graphs. The study culminates in a logical analysis showcasing that both modal and guarded fragments exhibit similar expressive capabilities when interpreting logics over undirected graphs.
Stats
In each iteration, vertices receive a message on each incoming edge.
The number of parameters is independent of the size of the input graph.
A query Q is uniformly expressible by a 2-GNN with SUM aggregation but not by a 1-GNN with SUM aggregation.
All queries non-uniformly expressible by families of 2-GNNs are also non-uniformly expressible by families of 1-GNNs.
All queries non-uniformly expressible by families of bounded depth and polynomial size are also non-uniformly expressible by families of bounded depth and polynomial size.
Quotes
"The question whether the two versions differ in their expressivity has been mostly overlooked in the GNN literature."
"By proving that modal and guarded fragment logics have similar expressivity over labelled undirected graphs..."
"The most natural is uniform expressivity... However, much literature considers non-uniform expressivity."
"In practical work, there seems to be a perception that 2-GNNs are superior..."
"Our focus here is on node classification or unary queries..."