AI Association Analysis Identifies Centrality Clusters Across Diverse Interaction Networks

A computational study published in Applied Network Science demonstrates that standard centrality indicators form tightly bounded, recurring clusters across diverse topological structures, allowing single representative measures to identify influential nodes without calculating redundant metrics. Researchers Bhawna Saxena, Adwitiya Sinha, Himansu Sekhar Pattanayak, Anokh Kishore, and Debasish Paul evaluated mathematical relationships among ten foundational centrality metrics across twelve empirical networks. By coupling pairwise Pearson correlation analysis with machine learning association rule mining, the investigative team established that complex structural measurements systematically co-occur according to predictable mathematical dependencies. The resulting framework confirms that individual metrics chosen from within identified clusters can reliably capture network-specific structural influence, offering an interpretable and computationally economical approach to large-scale network analysis across biological, technological, and social systems.

In network science, determining node prominence within interconnected structures has traditionally relied on an array of centrality formulations, each accentuating different structural characteristics. Measures such as degree centrality quantify immediate local connectivity, betweenness centrality tracks the frequency with which a node sits on shortest paths between other vertices, and closeness centrality evaluates positioning based on minimal geodesic distances across an entire community. While these mathematical tools offer valuable insights into information diffusion, virus tracking, and community cohesion, applying multiple overlapping formulations across massive graphs introduces severe computational bottlenecks, especially because the underlying statistical dependencies among diverse centrality indicators have remained insufficiently characterized across different network types.

The investigation was grounded in empirical evidence that statistical correlations observed among centrality indicators are driven primarily by the overarching topology of an evaluated network rather than representing fixed characteristics of the mathematical indicators themselves. Prior literature established that while metrics such as degree centrality frequently display high correlation with alternative indices, the degree of redundancy shifts noticeably across distinct connective architectures. Addressing this structural variability, the researchers sought to determine whether network instances belonging to the same functional class share consistent clusters of correlated indicators. Furthermore, the team investigated whether generic clusters persist irrespective of network classification, testing whether a unified rule base could systematically guide the selection of proxy indicators for finding influential nodes across diverse networked domains.

To capture broad topological variations, the experimental design evaluated twelve network instances organized into four representative classes: scientific collaboration networks, protein-protein interaction networks, online social trust networks, and peer-to-peer file-sharing systems. The scientific collaboration category encompassed three undirected co-authorship graphs from the arXiv repository and network science literature, covering General Relativity and Quantum Cosmology, High Energy Physics Theory, and Network Science. The biological class evaluated three undirected protein-protein interaction datasets, comprising one yeast interaction model and two distinct human binary interaction mappings. The online social network class examined three directed trust networks, including the Bitcoin OTC web of trust, the Bitcoin Alpha exchange platform, and the Advogato developer community. Finally, the peer-to-peer classification evaluated three directed file-sharing host topologies captured from the Gnutella network across distinct observation dates in August 2002. Across these twelve systems, graph sizes ranged from 1,461 to 9,877 nodes, average nodal degrees varied between 2.435 and 15.632, and average clustering coefficients spanned from 0.005 to 0.693, providing a structurally heterogeneous benchmark.

The analytical procedure extracted node-level measurements for ten baseline centrality indicators, including degree, closeness, eigenvector, betweenness, harmonic, subgraph, load, hub, authority, and Katz centralities. Because subgraph centrality requires complete network connectivity, it was excluded from directed networks characterized by disconnected components, leaving eight metrics for those graphs. For each network, the resulting node-by-metric matrix was transformed into a symmetric correlation matrix by computing pairwise Pearson correlation coefficients between indicator pairs. To isolate strongly associated relationships without arbitrary bias, the team established a filtering threshold based on the upper quartile value of the correlation distribution, formalized as the maximum of the third quartile or a baseline of 0.5. Values meeting or exceeding this threshold defined unweighted indicator graphs, which were subsequently decomposed into clusters using a maximal-clique partitioning technique that identified complete subgraphs of mutually correlated metrics while permitting individual indicators to reside in multiple clusters.

The structural validity of the derived clusters was tested by comparing the top fifty ranked nodes produced by each strongly correlated metric, confirming that high correlation coefficients corresponded directly to elevated counts of shared influential nodes. In the undirected collaboration networks, threshold values settled at 0.58 for General Relativity, 0.41 for High Energy Physics, and 0.52 for Network Science. The General Relativity and Network Science topologies yielded four identical maximal cliques: one pairing closeness with harmonic centrality, a second pairing betweenness with load centrality, a third grouping degree, eigenvector, subgraph, hub, and authority centralities, and an isolated single-metric cluster for Katz centrality.

Analysis of undirected protein-protein interaction networks revealed both preserved structural cores and divergence reflecting biological organization. For the yeast dataset, the upper quartile threshold reached 0.77, yielding four clusters that featured shared memberships for degree and Katz centralities across path-based and local groupings, while eigenvector, hub, and authority metrics clustered separately. In contrast, the two human protein interaction networks exhibited higher thresholds of 0.74 and 0.80 and condensed into three broader clusters: a large cohesive clique uniting degree, betweenness, load, hub, subgraph, authority, and eigenvector centralities; a distinct pairing of closeness and harmonic centralities; and an isolated Katz centrality cluster. Despite differences in edge distribution between yeast and human systems, every biological instance consistently retained the betweenness-load-degree association, the closeness-harmonic grouping, and the alignment of eigenvector, hub, and authority centralities, with Katz and subgraph centralities displaying weaker, network-specific associations.

In the directed online social networks and peer-to-peer file-sharing systems, directional edge flows introduced distinct structural alignments while preserving common clustering cores. Within the signed Bitcoin OTC and Bitcoin Alpha trust communities, where correlation thresholds reached 0.88 and 0.89, four cliques emerged, placing degree, betweenness, and load centralities in one group, degree, eigenvector, authority, and hub centralities in a second group, closeness and harmonic centralities in a third, and Katz centrality in isolation. In the Advogato developer network, this configuration remained stable, with hub centrality shifting into the path-based grouping. For the Gnutella peer-to-peer graphs, thresholds recorded at 0.69, 0.68, and 0.61 revealed four identical clusters across the first two collection dates, where degree, betweenness, and load grouped together, while degree, eigenvector, authority, and Katz centralities formed a second cluster, leaving hub and closeness-harmonic measures in independent groupings. The third Gnutella instance mirrored this behavior, omitting only the direct link between degree and eigenvector centrality due to an elevated graph density and higher average node degree.

Synthesizing results across all twelve network instances demonstrated that three core clusters consistently materialize regardless of domain or edge directivity: the local flow grouping of degree, betweenness, and load centralities; the spectral grouping of eigenvector and authority centralities, often joined by hub and subgraph metrics; and the neighborhood distance grouping of closeness and harmonic centralities. The physical drivers of these groupings trace directly to network topology. In scale-free environments characterized by prominent hubs, high-degree vertices dominate shortest paths, driving strong correlations among degree, betweenness, and load centralities. In small-world networks with tight local clustering, closeness and harmonic centralities converge because both track geodesic accessibility across compact neighborhoods. Conversely, in peer-to-peer topologies characterized by layered, hierarchical routing, Katz centrality frequently aligns with degree and eigenvector measures, whereas hub scores stand alone to designate role-specific terminal hosts.

To formalize these co-occurrences without relying solely on correlation thresholds, the authors implemented machine learning association rule mining, extracting fourteen significant dependency rules quantified through support, confidence, and lift metrics. In this analysis, antecedent and consequent support tracked the frequency of metric combinations across nodes, confidence measured conditional co-occurrence probabilities, and lift assessed whether combinations occurred more frequently than chance expectations. Several mined rules demonstrated deterministic relationships with confidence scores of 1.0, particularly configurations where simpler metrics implied recursive combinations involving eigenvector, authority, and hub scores. Crucially, the rule linking betweenness and degree centrality to load centrality yielded a maximum lift value of 4.7, identical to its inverse formulation, confirming that local path-traversal metrics reinforce one another at rates nearly five times higher than random occurrence. Similarly, multi-metric rules combining authority, hub, and eigenvector centralities to predict load centrality reached lift values of 3.5, whereas isolated metric pairs such as authority predicting eigenvector centrality exhibited modest lift values between 1.2 and 1.5, proving that multi-metric combinations provide substantially stronger predictive power than isolated indices.

The convergence of maximal-clique clustering and association rule mining confirms that redundant computation can be eliminated by choosing single representative proxies from established clusters. Network analysts evaluating large graphs can substitute degree centrality for load and betweenness centralities when measuring localized path importance, select eigenvector centrality as an efficient proxy for recursive prestige, and compute closeness centrality to reflect neighborhood reachability. The researchers noted that their current framework was established on unweighted and static graph representations under controlled structural parameters, leaving dynamic changes over time and weighted edge heterogeneity unexamined. Future work aims to incorporate temporal centrality formulations and edge weights into the association architecture, testing whether consistent centrality clusters persist across time-varying networks and evolving topological states.

Subject of Research: AI-driven association analysis and clustering of centrality indicators in complex networks

Article Title: Discovering centrality clusters in social and interaction networks using AI-driven association analysis

Article References: Saxena, B., Sinha, A., Pattanayak, H. S., Kishore, A., & Paul, D. (2026). Discovering centrality clusters in social and interaction networks using AI-driven association analysis. Applied Network Science, 11(1), Article 44. https://doi.org/10.1007/s41109-026-00794-3

Image Credits: AI Generated

DOI: 10.1007/s41109-026-00794-3

Keywords: network science, centrality indicators, degree centrality, betweenness centrality, closeness centrality, eigenvector centrality, association rule mining, maximal clique partitioning, collaboration networks, protein-protein interaction, online social networks, peer-to-peer networks, influential nodes, topological analysis, machine learning, Pearson correlation, Gnutella network, complex networks