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

Abstract

Abstract Information propagation over graph-structured data depends critically on the quality of node representations learned by graph neural networks (GNNs). However, how different GNN architectures organize embedding spaces and how these representations influence downstream propagation behavior remain insufficiently understood. This study presents a comprehensive empirical evaluation of ten representative GNN architectures on six attributed graph datasets using a unified unsupervised autoencoding framework and a fixed similarity-based propagation protocol. Rather than simulating a specific real-world diffusion process, the proposed framework provides a controlled setting for comparing how learned embeddings support propagation under identical conditions. Propagation behavior is quantified using normalized cascade size and multi-step activation dynamics. Temporal propagation patterns are analyzed using Dynamic Time Warping (DTW), while embedding organization is examined through principal component analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE), and clustering analysis. To investigate the role of graph topology, thirteen structural network metrics are analyzed using Pearson correlation, mixed-effects modeling, and SHapley Additive exPlanations (SHAP), followed by controlled graph intervention experiments. The results show that propagation behavior is jointly influenced by graph topology and the representations learned by different GNN architectures. PNA, GAT, GCN, and EvolveGCN consistently produce embedding spaces associated with larger propagation cascades across multiple datasets, although no single architecture is uniformly superior. The analyses further reveal architecture-specific structural fingerprints and demonstrate that identical topological perturbations can induce markedly different propagation responses. These findings provide new insights into how GNN representation learning shapes downstream similarity-based propagation behavior. The data and codes used in this study are available to reviewers at https://doi.org/10.5281/zenodo.17204159 . They will be made publicly available upon acceptance of the manuscript.