Caterpillar GNN: Replacing Message Passing with Efficient Aggregation

Fuente: arXiv
Saved in:
Bibliographic Details
Main Author: Černý, Marek
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914057817161728
author Černý, Marek
author_facet Černý, Marek
contents Message-passing graph neural networks (MPGNNs) dominate modern graph learning. Typical efforts enhance MPGNN's expressive power by enriching the adjacency-based aggregation. In contrast, we introduce an efficient aggregation over walk incidence-based matrices that are constructed to deliberately trade off some expressivity for stronger and more structured inductive bias. Our approach allows for seamless scaling between classical message-passing and simpler methods based on walks. We rigorously characterize the expressive power at each intermediate step using homomorphism counts over a hierarchy of generalized caterpillar graphs. Based on this foundation, we propose Caterpillar GNNs, whose robust graph-level aggregation successfully tackles a benchmark specifically designed to challenge MPGNNs. Moreover, we demonstrate that, on real-world datasets, Caterpillar GNNs achieve comparable predictive performance while significantly reducing the number of nodes in the hidden layers of the computational graph.
format Preprint
id arxiv_https___arxiv_org_abs_2506_06784
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Caterpillar GNN: Replacing Message Passing with Efficient Aggregation
Černý, Marek
Machine Learning
68T07 (Primary), 05C60, 05C85 (Secondary)
I.2.6; G.2.2
Message-passing graph neural networks (MPGNNs) dominate modern graph learning. Typical efforts enhance MPGNN's expressive power by enriching the adjacency-based aggregation. In contrast, we introduce an efficient aggregation over walk incidence-based matrices that are constructed to deliberately trade off some expressivity for stronger and more structured inductive bias. Our approach allows for seamless scaling between classical message-passing and simpler methods based on walks. We rigorously characterize the expressive power at each intermediate step using homomorphism counts over a hierarchy of generalized caterpillar graphs. Based on this foundation, we propose Caterpillar GNNs, whose robust graph-level aggregation successfully tackles a benchmark specifically designed to challenge MPGNNs. Moreover, we demonstrate that, on real-world datasets, Caterpillar GNNs achieve comparable predictive performance while significantly reducing the number of nodes in the hidden layers of the computational graph.
title Caterpillar GNN: Replacing Message Passing with Efficient Aggregation
topic Machine Learning
68T07 (Primary), 05C60, 05C85 (Secondary)
I.2.6; G.2.2
url https://arxiv.org/abs/2506.06784