How Betweenness, Closeness, and Eigenvector Algorithms Define Power in a Network

In January 2026, researchers publishing in arXiv proposed a unified framework to bring order to the more than 100 competing network centrality measures currently used to map everything from power grids to social influence. The preprint addresses a persistent mathematical ambiguity: defining what makes a single node important. In a graph of 10 million interconnected points, the most critical node is rarely the one with the most direct connections.[1]

The challenge of quantifying structural advantage dates back to sociologist Linton Freeman's 1977 formalization of betweenness, but modern computational graphs have stretched these definitions. According to Memgraph's technical documentation, "Betweenness centrality identifies the nodes that serve as bridges from one part of a graph to another." A node can have only two connections, yet if those two connections are the sole link between two massive, isolated clusters, that node controls the flow of the entire network.[3]

Measuring this control requires calculating the shortest paths between all possible pairs of nodes in the network. If a specific node lies on a high percentage of those shortest paths, its betweenness score rises. However, this exact calculation carries a time complexity of O(V³) for dense, unweighted graphs, meaning the computational time cubes as the number of vertices grows.[4]

For networks exceeding 100,000 nodes, exact betweenness becomes computationally prohibitive. NetworkX, a standard Python library for graph analysis, defaults to Brandes' algorithm, which reduces this complexity to O(V × E) for unweighted graphs. Even with this optimization, calculating betweenness on a 10-million node graph requires approximation techniques to return results in under three seconds.[3][4]

When the goal shifts from controlling flow to broadcasting information rapidly, closeness centrality provides a different mathematical lens. Closeness calculates the reciprocal of the sum of the length of the shortest paths between the node and all other nodes in the graph. A high closeness score indicates that a node can reach all other parts of the network in the fewest possible steps.[4][5]

This metric proves critical in epidemiology and logistics. In an April 2023 study published in MDPI analyzing plant networks across 54 species, researchers utilized closeness centrality to identify which biological nodes most efficiently distributed resources. The algorithm assumes that information or resources flow simultaneously along all shortest paths, making it ideal for modeling viral spread or fluid dynamics.[2]

Yet closeness struggles in highly fragmented networks. If a graph contains disconnected components, the distance between nodes in separate clusters is mathematically infinite, which zeroes out the closeness score for the entire network unless the algorithm is restricted to the largest connected component.[4]

Eigenvector centrality solves a different problem: the quality of connections rather than the sheer quantity. Proposed by Phillip Bonacich in 1987, this algorithm assigns relative scores to all nodes based on the principle that connections to high-scoring nodes contribute more to the score of the node in question than equal connections to low-scoring nodes.[1][5]

Google's original PageRank algorithm is a famous variant of eigenvector centrality, introducing a damping factor—typically set at 0.85—to model the probability that a random surfer continues clicking links rather than starting over. Neo4j's Graph Data Science library utilizes the power iteration method to calculate these scores at scale, updating each node's value based on its neighbors until the scores converge.[5]

The trade-off for eigenvector centrality is its sensitivity to localized echo chambers. A tightly knit cluster of nodes that only link to each other can artificially inflate their eigenvector scores, trapping the algorithm's mathematical weight in a closed loop.[1][5]

Choosing the correct algorithm dictates the outcome of the analysis. Degree centrality counts immediate neighbors, betweenness finds the bottlenecks, closeness identifies the most efficient broadcasters, and eigenvector maps the power brokers. Normalizing these scores allows data scientists to build composite rankings, but the raw metrics often disagree entirely on which node ranks first.[1][6]

The next frontier in network analysis moves beyond static topology. As graphs update in real-time—with edges forming and dissolving by the millisecond—algorithms must calculate temporal centrality without recomputing the entire network from scratch. The mathematical challenge is no longer just defining importance, but tracking it as it moves.[6]