Skip to content
Back to search results

Eigenvector Centrality

Introduced by
Bonacich, P.
Year
1972
Type
Centrality measure
Family
Network science
Font size

What does it do?

Eigenvector centrality, introduced by Bonacich in 1972, measures a node's influence by considering not just how many neighbors it has, but how influential those neighbors are.

A node scores highly if it is connected to other high-scoring nodes, making it a recursive, globally-aware measure of structural importance in a network.

The intuition behind the method

Being well-connected matters, but being connected to well-connected people matters even more. Eigenvector centrality captures this recursive idea: your score depends on your neighbors' scores, which in turn depend on their neighbors' scores, and so on throughout the whole network. It is as if influence flows continuously around the network until it settles at a stable equilibrium — the leading eigenvector of the adjacency matrix — where every node's score is proportional to the sum of its neighbors' scores.

History

Eigenvector centrality was formalized by Phillip Bonacich in a 1972 paper on factoring and weighting approaches to status scores, establishing the theoretical foundation for prestige-based centrality measures. The idea that a node's importance is proportional to the sum of its neighbors' importances predates Bonacich in the sociological notion of status, but his algebraic formalization made it computable. The measure gained broad adoption in social network analysis and later in bibliometrics and web graph analysis, directly inspiring Google's PageRank algorithm (1998), which adds a damping factor to handle directed and disconnected graphs.

When to use it

Use eigenvector centrality when you want to identify nodes that are influential not merely because they have many connections, but because they are embedded in densely connected, high-status regions of the network — typical in citation networks, social influence research, and organizational prestige analysis. It is appropriate for undirected or symmetric directed networks where the notion of mutual reinforcement makes substantive sense. Avoid it on disconnected graphs (scores become undefined for isolated components), on directed acyclic graphs (the leading eigenvector degenerates), or when local connectivity rather than global prestige is the research question. For directed networks with asymmetric influence, PageRank is generally preferable.

Limitations and considerations

  • Undefined or degenerate on disconnected graphs and directed acyclic graphs, where the leading eigenvector does not yield meaningful scores for all nodes.
  • Sensitive to network density and structure: in very sparse networks nearly all nodes collapse to near-zero scores except a small hub.
  • Does not account for path length or bridging roles; a node connecting two dense clusters may score low despite its structural importance.
  • For directed asymmetric networks the measure requires adaptation (e.g., PageRank), and naive application can be misleading.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Captures recursive, global influence rather than just local degree, revealing prestige and structural importance simultaneously.
  • Grounded in a principled linear-algebraic framework with a unique solution guaranteed by the Perron-Frobenius theorem for connected networks.
  • Naturally accounts for the quality of connections, not just their quantity, making it suitable for prestige and status research.
  • Efficient to compute via power iteration even on large sparse networks.
  • Directly interpretable: scores reflect proportional influence relative to the most central node.

Applications

  • Identifying prestige and influence in academic citation networks and co-authorship graphs.
  • Mapping power structures and opinion leaders in organizational and political networks.
  • Detecting hub nodes in biological protein-interaction and gene-regulatory networks.
  • Ranking web pages and documents by structural authority (foundational to PageRank).
  • Analyzing influence diffusion and key spreaders in social media networks.
  • Assessing status hierarchies in sociometric and friendship network studies.

How it works

How it works

1

The network is represented as an adjacency matrix A, where entry A_ij equals 1 (or the edge weight for weighted networks) if nodes i and j are connected, and 0 otherwise. This matrix encodes the full topology of connections on which centrality will be computed.

2

The largest eigenvalue and its associated eigenvector are computed from the adjacency matrix using power iteration or a standard linear algebra solver. The leading eigenvalue lambda_max is guaranteed to be real and positive for connected networks by the Perron-Frobenius theorem, ensuring a unique, interpretable solution.

Formula
\mathbf{A}\mathbf{x} = \lambda_{\max}\mathbf{x}
3

Each node is assigned the corresponding component of the leading eigenvector as its centrality score. The score satisfies the fixed-point condition: a node's score equals (1/lambda_max) times the sum of its neighbors' scores, so high-scoring nodes pull up the scores of their neighbors recursively.

Formula
x_i = \frac{1}{\lambda_{\max}} \sum_{j \in N(i)} x_j
4

Scores are typically scaled so that the maximum value equals 1, making them comparable across networks of different sizes. The relative ranking of nodes is preserved regardless of the normalization convention used.

Common pitfalls

  • Applying eigenvector centrality to a disconnected network without first extracting the largest connected component, leading to zero scores for entire subgraphs.
  • Conflating eigenvector centrality with degree centrality and failing to report which measure is used; they can yield very different rankings.
  • Using it on a directed network without verifying that the graph is strongly connected or adapting to a directed variant such as PageRank.
  • Treating small differences in normalized scores as meaningful when the network is sparse and scores are numerically compressed near zero.
  • Omitting sensitivity checks: centrality rankings can shift substantially if a few high-degree hub nodes are removed.

Frequently asked questions

How does eigenvector centrality differ from degree centrality?

Degree centrality counts only immediate neighbors, treating all connections as equal. Eigenvector centrality weights each neighbor by that neighbor's own centrality, so a connection to a highly central node contributes more than a connection to a peripheral one. A node with few but very influential neighbors can outrank a high-degree node connected mostly to peripheral nodes.

Why does eigenvector centrality fail on directed acyclic graphs?

In a directed acyclic graph there are no cycles, so influence cannot flow back upstream. The leading eigenvector of the adjacency matrix assigns zero scores to all nodes except those with no outgoing edges (sinks), making the measure degenerate. PageRank solves this by adding a teleportation term that redistributes weight uniformly.

Can I use eigenvector centrality on weighted networks?

Yes. Replace the binary adjacency matrix with a weighted adjacency matrix and apply the same eigen-decomposition. The leading eigenvector then reflects both the number and strength of connections. Ensure that edge weights are non-negative for the Perron-Frobenius guarantees to hold.

How should I report eigenvector centrality in a paper?

Report the normalized eigenvector centrality scores (max = 1), specify the software and version used (e.g., NetworkX, igraph), state whether the network was directed or undirected, and confirm that the largest connected component was used if the graph is disconnected. A visualization coloring or sizing nodes by their score greatly aids interpretation.

What is the relationship between eigenvector centrality and PageRank?

PageRank is a directed, damped generalization of eigenvector centrality. It adds a damping factor (typically 0.85) and a uniform teleportation probability to handle disconnected and directed graphs. For undirected connected networks, PageRank and eigenvector centrality produce nearly identical rankings.

Related methods

Sources

  1. Bonacich, P. (1972). Factoring and weighting approaches to status scores and clique identification. Journal of Mathematical Sociology, 2(1), 113–120.

    Open DOI reference
    Copied.
  2. Eigenvector centrality. Wikipedia.

    Open publication reference
    Copied.