Exploring the Influence of Graph Neural Network-Based Link Prediction on Social Contagion Dynamics

Social contagion shapes how information spreads through networks, influencing critical processes ranging from the dissemination of news and job opportunities to the propagation of social movements and potentially misleading content. The structure of social networks plays a pivotal role in determining the reach and impact of these diffusion processes, with network topology critically modulating how individuals access and interpret information. Within this context, link prediction (LP) emerges as a crucial mechanism for understanding and potentially manipulating network structures. The primary goal of LP is to determine whether two nodes in a network are likely to form a connection, thereby potentially reshaping the network's information transmission capabilities. While numerous LP methods have been proposed in the literature, along with various methodologies, biases research, and evaluation approaches, the relationship between LP and social diffusion processes remains less thoroughly explored, particularly concerning Graph Neural Network (GNN)-based LP algorithms. In this study, we systematically analyze four distinct GNN-based LP models to investigate how predicted network structures influence social contagion dynamics. Our research employs six diverse datasets, characterized by comprehensive node-level centrality measures and graph-level topological metrics. By leveraging this methodological variability, we aim to provide a nuanced understanding of how network characteristics correlate with social diffusion and how they are modulated by different LP models. We model social contagion using both simple and complex contagion frameworks through epidemic modeling techniques. Our findings reveal that LP models consistently reshape network structures in ways that significantly influence contagion dynamics. By introducing structural shortcuts or targeting hub nodes, these models enhance information diffusion, particularly in denser networks with high average degrees and clustering coefficients. Additionally, we observe that the impact of LP varies between simple and complex contagion processes, with attention-based models like Graph Transformers facilitating broader propagation and Graph Convolutional Networks (GCNs) forming localized clusters under specific conditions. Critically, measures such as Complex Path Centrality and node degree emerge as key predictors of contagion susceptibility, highlighting the intricate interplay between network topology and social diffusion behavior. This work contributes a comprehensive overview of GNN-based LP methods, network and node characterization, and social contagion modeling, taking an initial step toward bridging existing literature gaps and advancing the understanding of LP's impact on social dynamics.