Source-linked AI summary
The Diverse Club: The Integrative Core of Complex Networks
M. A. Bertolero, B. T. T. Yeo, M. D'Esposito
TL;DR
Networks often identify high-degree rich-club nodes as the basis of global integration, but the relative integrative roles of rich and diversely connected nodes remain unclear. This paper analyzes diverse and rich clubs across biological and engineered networks and models their evolutionary emergence, finding that the diverse club more consistently exhibits properties of an integrative core.
Problem
The integrative roles of high-strength rich-club nodes versus high-participation diverse-club nodes remain insufficiently distinguished across complex networks.
Source: Discussion · p. 13Method
The authors compare club connectivity and network-integration properties across five network types and simulate selection balancing modularity with efficient integration.
Source: Results · pp. 4–5, Discussion · p. 14Results
The diverse club is more interconnected than the rich club in every analyzed network, and removing its edges more strongly reduces efficient global communication.
Source: Discussion · pp. 13–14Takeaways & Limitations
The diverse club, rather than the rich club, is the proposed integrative core of complex networks and may reflect distinct evolutionary pressures.
Source: Discussion · pp. 14–15Community-detection algorithms have resolution limits, and networks may contain multiple valid community scales, making community structure scale-dependent.
Source: Supplementary Methods · p. 17
Abstract
from arXiv · showhide
A complex system can be represented and analyzed as a network, where nodes represent the units of the network and edges represent connections between those units. For example, a brain network represents neurons as nodes and axons between neurons as edges. In many networks, some nodes have a disproportionately high number of edges. These nodes also have many edges between each other, and are referred to as the rich club. In many different networks, the nodes of this club are assumed to support global network integration. However, another set of nodes potentially exhibits a connectivity structure that is more advantageous to global network integration. Here, in a myriad of different biological and man-made networks, we discover the diverse club--a set of nodes that have edges diversely distributed across the network. The diverse club exhibits, to a greater extent than the rich club, properties consistent with an integrative network function--these nodes are more highly interconnected and their edges are more critical for efficient global integration. Moreover, we present a generative evolutionary network model that produces networks with a diverse club but not a rich club, thus demonstrating that these two clubs potentially evolved via distinct selection pressures. Given the variety of different networks that we analyzed--the c. elegans, the macaque brain, the human brain, the United States power grid, and global air traffic--the diverse club appears to be ubiquitous in complex networks. These results warrant the distinction and analysis of two critical clubs of nodes in all complex systems.
Results
Across biological and man-made networks, the diverse club was more interconnected, more broadly distributed across communities, and more critical for efficient global communication than the rich club. A generative model selected for modularity and efficiency produced a highly interconnected diverse club but not a rich club.
The diverse club exhibits stronger clubness than the rich club
- The diverse club was typically at least as interconnected as the rich club, and its edges had higher edge betweenness in almost all networks.Node betweenness was not consistently different, but edge betweenness was significantly higher for the diverse club across algorithms in almost all networks.Source: Results · pp. 6, 10, Figure 2 · p. 7
The rich club and the diverse club are mostly non-overlapping clubs that exhibit different network properties
- The diverse and rich clubs were predominantly composed of different nodes, while the diverse club appeared centrally located and represented communities at least as broadly as the rich club.Human-network overlap was no more than 23%, whereas overlap ranged from 6–76% in functional c. elegans networks and 18–63% in structural networks.Source: Figure 1 · p. 5, Results · p. 8, Figure 3 · p. 9
Targeted attacks of intra-diverse-club connections decrease efficient integration
- Removing intra-club edges increased shortest-path sums more for the diverse club than the rich club in every network, indicating greater importance for efficient global communication.Edges between diverse club nodes were removed in 10,000 iterations, with 50–90% of edges removed while avoiding graph disconnection.Source: Results · p. 10
Diverse club, but not rich club activity, increases as more communities are engaged by cognitive tasks
- The diverse club’s activity increased as tasks engaged more communities or cognitive components, whereas rich-club activity decreased.For the diverse club, mean correlations were r=0.45 with communities and r=0.395 with cognitive components; corresponding rich-club correlations were r=-0.37 and r=-0.45.Source: Results · pp. 10–11
The diverse club potentially evolved as a solution to efficient integration in modular networks
- At a selection ratio of 0.75 for modularity Q to 0.25 for efficiency, the generative model achieved high modularity, efficiency, and diverse-club clubness while matching brain-network distributions.Across 1000 model and 1000 random-model iterations, the model’s diverse-club clubness exceeded the random model’s at higher ranks, while the rich club did not show high clubness.Source: Figure 4 · p. 12, Results · p. 13
Discussion
The diverse club, rather than the rich club, shows the strongest evidence of an integrative network core across biological and man-made networks. The clubs may have distinct functions and evolutionary origins, with the diverse club supporting efficient integration and the rich club potentially stabilizing network dynamics.
Comparative club properties
- The diverse club is more interconnected than the rich club in every analyzed network, reaching up to fourfold greater interconnectedness in the human brain, while few nodes overlap.Networks included seven human brain states, C. elegans, macaque brain, the United States power grid, and global air traffic.Source: Discussion · p. 13
Integrative function
- Diverse-club edges are more critical than rich-club edges for efficient global communication: removing them increases summed shortest paths significantly more across all networks.Diverse-club nodes also span more communities and have higher edge betweenness, despite equal node betweenness centrality.Source: Discussion · pp. 13–14
- In humans, diverse-club activity increases as cognitive tasks engage more communities, unlike rich-club activity, consistent with greater integration across network communities.This activity difference emerges as task complexity increases.Source: Discussion · p. 14
Evolutionary origins
- Selecting for a balance between modularity and efficient integration produces a highly interconnected diverse club but not a rich club, with higher diverse-club clubness than a random null model.This supports distinct evolutionary pressures favoring the diverse club’s combination of modularity and efficient integration.Source: Discussion · p. 14
Functional interpretation
- These findings motivate analyzing both clubs and identify the diverse club, rather than the rich club, as the likely integrative core of complex networks.The rich club may instead help stabilize spontaneous network dynamics through high in-degree and slower-timescale integration.Source: Discussion · p. 15
- Human rich-club anatomy and cognition support a possible stability role: default-mode nodes are distant from primary sensory and motor networks and support stimulus-independent processes such as daydreaming.This proposed slow-processing distinction from the diverse club remains unresolved and warrants further investigation.Source: Discussion · p. 15
Supplementary Methods
The supplementary methods define network measures and a generative evolutionary model, applying them across biological and man-made networks. Supplementary analyses test clubness under alternative methodological specifications and compare diverse and rich clubs across network measures, cognitive-task engagement, and model parameters.
Community detection
- Community structure was detected using modularity-maximizing, spectral, label-propagation, edge-betweenness, and InfoMap algorithms, each embodying a different community definition [52–56, 59].Modularity Q compares within-community edge weights with those expected in a randomized null model preserving node strengths; label propagation uses neighbor labels, edge betweenness removes high-betweenness edges, and InfoMap models information flow.Source: Supplementary Methods · pp. 15–17
- Because community algorithms have resolution limits and networks support multiple valid scales, the analysis used stability-based Louvain and dendrogram-based Walktrap to specify community scale without thresholding dense networks [60, 61].Stability uses the time-scale of a dynamical process as an intrinsic resolution parameter, whereas Walktrap permits cutting its dendrogram at a desired number of communities.Source: Supplementary Methods · p. 17
Participation and clubness
- Participation coefficients quantify how evenly each node’s edge weights span communities, while clubness ranks nodes by strength or participation and normalizes within-club connectivity against degree- and strength-preserving random networks.Participation is maximal when edge-weight sums are equal across communities and zero when all edges reach one community; normalized clubness is additionally standardized by the random-graph standard deviation.Source: Supplementary Methods · pp. 17–18
Efficiency
- Network efficiency was defined as the inverse of the summed shortest-path distances, with binary edges used for shortest-path calculations across all networks and analyses.Weighted shortest paths were excluded because minimizing summed edge weights represents short travel routes, which was considered inappropriate for these networks.Source: Supplementary Methods · p. 18
Network data
- Analyses covered functional and structural C. elegans, human fMRI, macaque cortical, global air-traffic, and United States power-grid networks using network-specific node, edge, thresholding, and preprocessing procedures.Functional C. elegans and human networks were thresholded across costs, whereas macaque, air-traffic, and power-grid networks were analyzed without thresholding; human data included 471 subjects and six tasks plus rest, excluding emotion.Source: Supplementary Methods · pp. 18–19
Generative evolutionary model
- The generative model repeatedly removed edges favoring higher modularity and shorter paths, randomly rewired them at constant density, and evaluated similarity to human networks using inverse Kullback–Leibler divergence.It began with 100 randomly connected nodes at 5% density, removed and rewired 5% of edges per iteration for 150 iterations, and compared against a random-edge null model.Source: Supplementary Methods · p. 20
Robustness analyses
- Participation coefficients were compared across community-detection algorithms and runs for flight-traffic, macaque, and US power-grid networks.Supplementary analyses also evaluated clubness across these network classes under multiple null-model specifications.Source: Figure 35 · p. 55, Figure 36 · p. 56, Figure 37 · p. 57
Clubness calculations
- Clubness was recalculated using degree-preserving random graphs, with variants retaining node strength and shuffling edge weights among nodes with the same degree.Additional analyses normalized clubness by the standard deviation across random graphs.Source: Figure 38 · p. 58, Figure 39 · p. 59, Figure 40 · p. 60
Club properties
- Supplementary comparisons examined club overlap, community span, node betweenness, and within-club edge betweenness for diverse and rich clubs across functional and structural networks.Community analyses quantified shared club membership and differences in the percentages of communities containing each club; centrality tests used Bonferroni correction.Source: Figure 50 · p. 70, Figure 51 · p. 71, Figure 52 · p. 72
Targeted attacks
- Targeted attacks removed 50–90% of club edges across 10,000 iterations while skipping removals that disconnected the graph, then measured increases in summed shortest paths.The analyses compared attacks on rich and diverse clubs in functional and structural networks, with Bonferroni-corrected P values.Source: Figure 59 · p. 79, Figure 60 · p. 80, Supplementary Methods · p. 81
Cognitive and generative analyses
- Additional analyses correlated cognitive-task community or component engagement with diverse- and rich-club activity and tested alternative Q-to-efficiency ratios in generative models.The generative-model analyses included Q-to-efficiency ratios of 0.70 and 0.80.Source: Figure 62 · p. 82, Figure 63 · p. 83, Figure 64 · p. 84