A universal assortativity measure for network analysis

Characterizing the connectivity tendency of a network is a fundamental problem in network science. The traditional and well-known assortativity coefficient is calculated on a per-network basis, which is of little use to partial connection tendency of a network. This paper proposes a universal assortativity coefficient(UAC), which is based on the unambiguous definition of each individual edge’s contribution to the global assortativity coefficient (GAC). It is able to reveal the connection tendency of microscopic, mesoscopic, macroscopic structures and any given part of a network. Applying UAC to real world networks, we find that, contrary to the popular expectation, most networks (notably the AS-level Internet topology) have markedly more assortative edges/nodes than dissortaive ones despite their global dissortativity. Consequently, networks can be categorized along two dimensions–single global assortativity and local assortativity statistics. Detailed anatomy of the AS-level Internet topology further illustrates how UAC can be used to decipher the hidden patterns of connection tendencies on different scales.

💡 Research Summary

The paper addresses a fundamental limitation of the classic assortativity coefficient introduced by Newman, which quantifies the tendency of nodes with similar degree to connect but does so only at the whole‑network level. The authors propose a “Universal Assortativity Coefficient” (UAC) that decomposes the global assortativity into contributions from individual edges, thereby enabling the measurement of assortative or disassortative tendencies on any sub‑graph, node, community, or set of inter‑group links.

Starting from the definition of the remaining‑degree distribution q(k) and its mean U_q, the authors rewrite Newman’s r as a normalized covariance between the remaining degrees at the two ends of a randomly chosen edge. This leads to a simple expression for the contribution of a single edge e with endpoint remaining degrees j and k:

 ρ_e = (j – U_q)(k – U_q) / (M·σ_q²)

where M is the total number of edges and σ_q² is the variance of q(k). Positive ρ_e denotes an assortative edge, negative ρ_e a disassortative edge, and the absolute value |ρ_e| quantifies the strength of that contribution. In the degenerate case of a perfectly regular graph (σ_q = 0) the authors set ρ_e = 1/M for all edges, preserving the property that the sum of all ρ_e equals the global r.

The Universal Assortativity Coefficient for any target edge set E_target is then defined as the sum of the corresponding ρ_e values:

 ρ(E_target) = Σ_{e∈E_target} ρ_e

When E_target = E (the whole edge set) the expression reduces to Newman’s r, confirming that UAC is a true generalization. By choosing different target sets, one can compute node‑level assortativity (ρ_v = Σ_{e incident to v} ρ_e), community‑level assortativity (edges internal to a community), or inter‑community assortativity (edges crossing two communities).

The authors apply UAC to a diverse collection of real‑world networks: technical (AS‑level Internet, router topology, airline network), biological (protein‑protein interaction, C. elegans neural network, food webs), social (collaboration networks, citation networks), online social (Epinions), email (Enron), and a synthetic Erdős–Rényi graph. Their findings overturn the common intuition that a globally disassortative network must be dominated by disassortative edges. For many networks with negative global r (e.g., AS‑2011‑6, Router, USAir), the proportion of assortative edges P(ρ_e>0) exceeds 50 %, sometimes reaching 60 %. The overall disassortativity is instead driven by a smaller number of strongly disassortative edges whose average strength |ρ_e| is larger than that of the majority of assortative edges. Conversely, networks that are globally assortative (e.g., SCN, CA‑HepTh, CA‑GrOc) show both a higher proportion of assortative edges and higher average strength, reinforcing the global trend.

A particularly insightful case study is the AS‑level Internet topology. The authors partition the ASes first by Regional Internet Registries (RIRs) and then by country/region, creating a two‑level hierarchy. Using UAC they compute intra‑RIR and inter‑RIR assortativity coefficients. Most RIRs are internally assortative, but ARIN and RIPE‑NCC display disassortative connections to other RIRs, reflecting their central role in the global Internet backbone. At the country level, most nations are internally assortative, while a few (US, CA, GB, EU, DE) tend to connect disassortatively with the rest of the world. The heat‑map of inter‑country assortativity reveals a clear diagonal of internal assortativity and off‑diagonal patches of disassortativity, highlighting geopolitical and infrastructural influences on connectivity.

Based on the joint distribution of global r (positive or negative) and the local statistic P(ρ_e>0) (greater or less than 50 %), the authors propose a four‑quadrant classification of networks:

1. Globally assortative, majority assortative edges.
2. Globally assortative, majority disassortative edges.
3. Globally disassortative, majority disassortative edges.
4. Globally disassortative, majority assortative edges.

This taxonomy provides a nuanced view of network mixing patterns and suggests that many real systems fall into the fourth quadrant—globally disassortative yet locally dominated by assortative edges—an observation that would be invisible without edge‑level analysis.

In summary, the paper introduces a mathematically sound, computationally straightforward extension of assortativity that operates at any desired scale. By attributing a clear, signed contribution to each edge, UAC enables researchers to dissect the structural drivers of global mixing patterns, to relate them to dynamical processes (e.g., epidemic spreading, robustness to attacks), and to uncover hidden hierarchical or geographic organization in complex networks. The extensive empirical validation across multiple domains demonstrates both the practicality and the explanatory power of the proposed measure.