Graph neural networks shape propagation differently, large comparison study finds

Graph neural networks, the machine learning models that learn from network-structured data, do not all organize what they learn in the same way, and those differences ripple into how information spreads across a graph. That is the central finding of a new empirical study published in Applied Network Science, which compares ten representative graph neural network architectures on six attributed graph datasets under a single, controlled experimental protocol.

The study, authored by Amin Mahmoudi and Dariusz Jemielniak, was published on 7 October 2026 as an open access research article. Its stated motivation is a gap in understanding: how different graph neural network architectures organize their embedding spaces, and how those learned representations influence downstream propagation behavior, remain insufficiently understood.

To close that gap, the researchers evaluated all ten architectures using a unified unsupervised autoencoding framework and a fixed similarity-based propagation protocol. Importantly, the authors emphasize that the framework is not intended to simulate any specific real-world diffusion process. Instead, it provides a controlled setting in which the question is how well the embeddings learned by each architecture support propagation when every other condition is held identical.

Propagation behavior was quantified in two ways: normalized cascade size, which captures how far activations spread, and multi-step activation dynamics, which track how activation unfolds over successive steps. To examine temporal patterns in propagation, the team applied Dynamic Time Warping. Embedding organization was examined through principal component analysis, t-distributed stochastic neighbor embedding, and clustering analysis.

Because graph topology could plausibly explain much of the observed behavior, the study also analyzed thirteen structural network metrics. These were examined using Pearson correlation, mixed-effects modeling, and SHAP, a feature-attribution method, followed by controlled graph intervention experiments in which the researchers deliberately altered graph structure to observe the consequences.

The results show that propagation behavior is jointly influenced by graph topology and by the representations learned by different architectures. Four architectures in particular, PNA, GAT, GCN, and EvolveGCN, consistently produced embedding spaces associated with larger propagation cascades across multiple datasets. The authors are careful to note, however, that no single architecture was uniformly superior across all settings.

Beyond the rankings, the analyses revealed architecture-specific structural fingerprints, meaning that each architecture leaves a characteristic signature in how its embeddings relate to graph structure. The intervention experiments added a further twist: identical topological perturbations could induce markedly different propagation responses depending on the architecture involved. In other words, changing the same edge or node in the same way does not produce the same downstream effect across models.

The findings offer new insight into how representation learning in graph neural networks shapes downstream, similarity-based propagation. The study also underscores the value of controlled comparison frameworks, since differences in propagation behavior emerged under identical conditions rather than from differences in task setup.

The authors report that the data and code used in the study are available to reviewers via a Zenodo repository, and that they will be made publicly available upon acceptance of the manuscript. The published version is currently shared as an accepted, citable article that remains subject to further editorial edits before the final Version of Record replaces it.

Subject of Research: How graph neural network architectures organize embeddings and influence similarity-based information propagation across graph datasets

Article Title: A comparative analysis of graph neural networks through similarity-based information propagation across diverse network structures

Article References: Mahmoudi, A., & Jemielniak, D. (2026). A comparative analysis of graph neural networks through similarity-based information propagation across diverse network structures. Applied Network Science. https://doi.org/10.1007/s41109-026-00819-x

Image Credits: AI Generated

DOI: 10.1007/s41109-026-00819-x

Keywords: graph neural networks, information propagation, embedding spaces, network topology, GAT, GCN, EvolveGCN, PNA, SHAP analysis, cascade size, Applied Network Science, machine learning