What Eigenvector Centrality Is
Eigenvector centrality is a network-science measure that scores each node by the principle that connections to important nodes count more. Rather than simply counting links (degree centrality), it weighs each connection by the importance of the neighbor it leads to.
It is one of the oldest recursive centrality measures — and the direct conceptual ancestor of PageRank, which is best understood as eigenvector centrality modified for the directed, noisy structure of the web.
What Eigenvector Centrality Ranks
The measure ranks nodes in a network — web pages, people in social graphs, papers in citation networks, proteins in interaction maps — by their position within the connection structure.
Each node receives a score; the highest-scoring nodes are those plugged into the network’s most important regions.
Core Inputs Used by Eigenvector Centrality
Exactly one input:
- The network’s adjacency structure — which nodes connect to which (optionally weighted)
No content, attributes, or external quality signals are used. Importance is read entirely from topology.
How Eigenvector Centrality Is Calculated (High-Level)
The recursive definition becomes linear algebra:
- Write the rule: each node’s score equals the (scaled) sum of its neighbors’ scores.
- Collected across all nodes, this is the eigenvector equation Ax = λx — the score vector must be an eigenvector of the adjacency matrix.
- The meaningful solution is the dominant eigenvector (largest eigenvalue), whose entries are non-negative by the Perron–Frobenius theorem when the network is connected.
- In practice, scores are found by power iteration: start with equal scores, repeatedly recompute each node as the sum of its neighbors, normalize, and repeat until stable — the same computational trick PageRank later used.
Conceptual model: Ask every node “who are your friends?”, let importance flow inward along each answer, and repeat until the flow stops changing.
Key Parameters or Factors
- Directed vs. undirected — in directed networks the measure uses incoming links and can leave entire regions scoreless (the problem PageRank’s damping fixes)
- Weights — edges can carry strengths, making important connections count more
- Convergence conditions — guaranteed on connected, non-bipartite graphs; real networks often need adjustments
Update Frequency
Centrality is recomputed when the network changes — there is no standing schedule; it is a property of a given graph snapshot.
Known Limitations and Criticisms
- Concentration — in many real networks nearly all centrality pools in a few nodes, leaving the rest indistinguishable
- Directed-network degeneracy — nodes with no incoming links get zero, and entire downstream regions can collapse (the “dangling node” problem)
- No topic sense — like PageRank, the measure is blind to what nodes are about
- Static snapshot — says nothing about how importance evolves
Where Eigenvector Centrality Is Used
The measure is used in:
- Social network analysis — identifying influential actors
- Web and citation analysis — as the foundation beneath PageRank and HITS
- Biology — ranking proteins and genes in interaction networks
- Infrastructure and epidemiology — finding critical and high-spread nodes
Summary
Eigenvector centrality is the purest form of recursive ranking: you matter if the nodes that point at you matter. Nearly every limitation engineers later patched — damping, personalization, hub/authority splits — was a response to this clean idea meeting messy real networks. Understanding it is the key that unlocks the whole link-analysis family.
References and Sources
- Bonacich, P. Power and Centrality: A Family of Measures (1987).
- Newman, M. E. J. Networks: An Introduction.
- Wikipedia. Eigenvector centrality.