Spectral centrality measures in complex networks

Complex networks are characterized by heterogeneous distributions of the degree of nodes, which produce a large diversification of the roles of the nodes within the network. Several centrality measures have been introduced to rank nodes based on their topological importance within a graph. Here we review and compare centrality measures based on spectral properties of graph matrices. We shall focus on PageRank, eigenvector centrality and the hub/authority scores of HITS. We derive simple relations between the measures and the (in)degree of the nodes, in some limits. We also compare the rankings obtained with different centrality measures.

💡 Research Summary

This paper provides a comprehensive review and comparative analysis of three spectral centrality measures—PageRank, eigenvector centrality, and the hub/authority scores of the HITS algorithm—within the context of complex networks. After introducing the basic adjacency matrix representation for directed and undirected graphs, the authors first formalize PageRank (PR) as the principal eigenvector of a stochastic transition matrix M that blends a random walk along outgoing links with a uniform “teleportation” term controlled by the damping factor q. They examine the two extreme regimes q→0 (pure random walk) and q→1 (complete teleportation). In the small‑q limit, they show analytically that PR values become integer multiples of q/n, where n is the number of nodes, and that each node’s PR is simply the count of its predecessors (nodes that can reach it via directed paths). Using a tree subgraph as an illustrative example, they derive a master equation for the distribution P_PR(l) of PR “levels” l in the Dorogovtsev‑Mendes‑Samukhin (DMS) preferential‑attachment model (with out‑degree m=1). Solving the master equation yields a power‑law tail P_PR(l)∼l⁻², a result confirmed by large‑scale simulations (10⁶ nodes) and shown to hold for other growth models such as Barabási‑Albert and copying models.

The second measure, eigenvector centrality (EV), is defined as the leading eigenvector of the transpose adjacency matrix Aᵀ. The authors note that nodes with zero indegree obtain zero centrality, which can be problematic in sparse real‑world graphs. To remedy this, they introduce a regularized version α‑EV: x = α Aᵀx + ε, where ε is a small uniform prestige term and α controls the relative weight of neighbor contributions. This formulation guarantees non‑zero scores for all nodes and, after normalizing by the sum of scores, yields a probability‑like distribution.

The third family, HITS, assigns each node a hub score x and an authority score y that satisfy the coupled equations y = Aᵀx and x = Ay. Substituting leads to eigenvalue problems for the symmetric matrices AAᵀ (for hubs) and AᵀA (for authorities). The authors emphasize that, unlike PR and EV, HITS explicitly captures a bipartite structure: hubs point to authorities and authorities are cited by hubs. Consequently, hub scores correlate with out‑degree, while authority scores correlate with indegree.

Having established the theoretical foundations, the paper proceeds to empirical comparisons on several real networks, including a web graph, an Internet topology, and a citation network. Correlation analyses reveal that PR and EV rankings are highly similar, especially when q is small, reflecting their shared dependence on indegree and the prestige of neighbors. HITS, however, produces distinct rankings that highlight nodes playing a bridging role between high‑out‑degree hubs and high‑in‑degree authorities. The authors illustrate that hub scores concentrate on nodes with large out‑degree, whereas authority scores concentrate on nodes with large indegree, confirming the algorithm’s suitability for tasks such as identifying authoritative content or influential spreaders.

In conclusion, the study demonstrates that spectral centralities provide richer information than simple degree‑based metrics. PageRank captures stochastic flow dynamics, eigenvector centrality reflects global connectivity patterns, and HITS uncovers a two‑mode structure useful for domains where citation or endorsement relationships matter. The paper also outlines avenues for future work, such as extending these measures to temporal, multilayer, or non‑linear diffusion contexts, thereby underscoring the versatility and limitations of spectral centrality in complex network analysis.