Comparing gravity-based centralities across divergent network topologies: evidence from internet AS and citation DAGs

We compare eleven gravity-based centrality models across two divergent Directed Acyclic Graph (DAG) domains: Internet Autonomous System Networks (ASNs) and scientific citation networks. Using Kendall’s τ , Spearman’s ρ, and monotonicity metrics, we evaluate ranking resolution, benchmark comparison, and structural distinctiveness. Results show that gravity-based models achieve strong ranking resolution once leaf nodes are removed, and several variants remain robust even with leaves included in citation networks. The DKE, Entropy-Based, K-Shell Based, and Multi-Characteristics models perform best, while the Extended and Laplacian variants underperform. Closeness consistently emerges as the strongest classical baseline but remains structurally independent from gravity-based families. Correlation analyses reveal three distinct clusters of gravity-based models, indicating that different formulations capture complementary structural properties. We show that gravity-based models offer a scalable and adaptable framework for DAG analysis and motivate new gravity formulations that build on the discriminative strengths of Closeness.