Degree Centrality in Network Analysis

• Degree centrality is a foundational network measure defined as the number of a node's direct neighbors, identifying influential hubs.
• Its computational simplicity (O(N+L)) makes it ideal for analyzing large-scale networks and establishing baseline metrics.
• Extensions like generalized degree and m-order degree mass integrate local and global network effects for advanced applications.

Degree centrality is the most fundamental local measure of node prominence in network science. For a simple undirected network, it is defined as the number of immediate neighbors of a node, or equivalently, the sum of the entries in the node's corresponding row (or column) of the adjacency matrix. Degree centrality underpins a broad array of applications, forms the basis of several generalizations and extensions, and maintains a central place in both theoretical and applied network analysis.

1. Mathematical Formulation and Definitions

Let $G = (V, E)$ be a simple, undirected graph with $n = |V|$ nodes. The degree centrality $C_D(i)$ of node $i$ is: $C_D(i) = d(i) = \sum_{j \in V} a_{ij}$ where $a_{ij}$ are the entries of the adjacency matrix $A$. The normalized form,

$$C_D^{\mathrm{norm}}(i) = \frac{d(i)}{n-1} \in [0,1]$$

allows direct comparability across networks of different sizes (Ibnoulouafi et al., 2023). In directed graphs, one distinguishes in-degree and out-degree centralities.

Degree centrality admits a vector form for all nodes as $(A\mathbf{u})_i$, where $\mathbf{u}$ is the all-ones vector (Li et al., 2014).

2. Interpretations and Theoretical Foundation

Degree centrality operationalizes the notion of node "activity" or "visibility", identifying hubs as nodes with large numbers of direct links. Sciarra et al. derive degree centrality as the solution to a rank-1 least-squares estimator of the adjacency matrix: $$A_{ij} = \hat{A}_{ij} + \epsilon_{ij} = f(x_i, x_j) + \epsilon_{ij}$$ Taking $$f(x_i, x_j) = \frac{K_{\rm tot}}{N}(x_i + x_j - 1/N)$$, the optimum node score is proportional to degree (Sciarra et al., 2018). The unique contribution of a node is quadratic in its degree, reflecting that hubs disproportionately influence adjacency reconstruction.

Despite being severely information-compressive (reducing $N^2$ data to $N$ scores), degree centrality explains as much variance as any other single-component metric in large surjective reconstructions. Multi-component (e.g., SVD-based) generalizations can overcome this compression, capturing higher-order structural roles (Sciarra et al., 2018).

3. Extensions and Generalizations

Several nontrivial generalizations enhance or refine degree centrality:

Generalized Degree Centrality: Defined as the unique solution to

$(I + \varepsilon L)\mathbf{x} = \mathbf{d}$

where $L$ is the graph Laplacian and $\varepsilon > 0$ is a tunable parameter. This extension redistributes degree mass through network structure, with $\varepsilon$ governing the balance between purely local and globally diffused influence. For small $\varepsilon$, it reduces to degree; as $\varepsilon \rightarrow \infty$, it approaches uniformity. The method resolves degree ties by accounting for the connectivity of neighbors and remains within the range $[\min d_i, \max d_i]$. Iterative computation is efficient; rank monotonicity is guaranteed for sufficiently small $\varepsilon$ (Csató, 2015).

Degree Mass (m-order): The $m$-order degree mass $D^{(m)}(v)$ sums weighted degrees across the $m$-hop neighborhood: $D^{(m)}(v) = \sum_{k=1}^{m+1}[A^k\mathbf{u}]_v$ This forms a bridge between local and global centralities; for $m = 0$ it recovers degree, and as $m$ increases, it correlates increasingly with centralities such as eigenvector centrality (Li et al., 2014).

Cross-Layer Degree Centrality: For multi-layered networks $\text{MSN} = (V, E, L)$, degree centrality generalizes to aggregated cross-layer measures,

$$\text{CLDC}(x, a) = \frac{\sum_{\ell \in L} \sum_{y \in MN(x,a)} \left[w(x,y,\ell) + w(y,x,\ell)\right]}{(n-1)|L|}$$

where $MN(x,a)$ collects neighbors adjacent on at least $a$ distinct layers. This allows context-dependent weighting of multiplexity in real multirelational systems (Bródka et al., 2012).

4. Computational Efficiency and Scalability

Degree centrality is computationally minimal. For a network with $N$ nodes and $L$ edges, all degrees can be computed in $O(N+L)$ time, or $O(L)$ if only an edge list is stored (Li et al., 2014). This computational advantage underlies its use in massive networks and as a building block or baseline for more complex metrics.

Further, degree-based ranking can be estimated with high precision in very large graphs using sublinear random-walk sampling, exploiting inferred power-law degree distributions. For instance, given only a node’s observed degree and estimated global parameters (minimum, average, and maximum degree, as well as inferred exponent), one can tightly estimate degree-rank via

$R_p(u) = a k_u^{1-\gamma} + b$

with all parameters efficiently computable from a sublinear sample (Saxena et al., 2015, Saxena et al., 2015). Simulations on Barabási–Albert networks yield errors under 0.16% of network size, particularly for hub nodes.

5. Empirical Observations and Comparative Analysis

Degree centrality correlates strongly with several global metrics:

• Betweenness: Pearson coefficient $\rho > 0.95$ in synthetic random graphs; often $>0.6$ in diverse real-world networks.
• Closeness: $\rho$ between closeness and 1st-order degree mass can exceed 0.95; with raw degree, correlations are also high except in networks with extreme structural heterogeneity (Li et al., 2014).
• Eigenvector centrality: Strong correlation with 2nd-order degree mass; $\rho$ up to 0.8–0.95 depending on network type.

In practice, simple degree-based selections are nearly optimal for control tasks such as maximizing or disrupting opinion spread (e.g., in the ICO model), with no complex global centrality outperforming degree or leverage-based strategies in phase-transition thresholds and largest-component size convergence (Li et al., 2014).

6. Functional Relationships and Redundancy with Other Centralities

Recent work has established explicit mathematical relations between degree and more complex centralities. Evans and Chen demonstrate that inverse closeness (farness) is, generically, an affine function of the logarithm of degree: $$\frac{1}{C_i} \approx \beta - \frac{1}{\ln z} \ln d_i$$ for appropriately fitted constants $\beta$ and $z$, where $z$ encodes the typical branching ratio for spheres around a node (Evans et al., 2021). In most real and model networks, closeness centrality thus contributes little additional information beyond degree, as confirmed by high Pearson correlations and small residual variance.

7. Limitations and Beyond-Degree Approaches

Purely local degree centrality often fails to differentiate nodes in nuanced structural contexts:

• Nodes with identical degrees may occupy core vs. periphery, or act as bridges between modules versus being embedded in trees (Ibnoulouafi et al., 2023).
• In multi-layered contexts, simple degree cannot model multiplex or functionally differentiated relations; sophisticated cross-layer definitions are required (Bródka et al., 2012).
• In dynamic or growing networks, early degree alone poorly predicts future influence or centrality. The Node Prominence Profile (NPP), combining degree with triadic closure statistics, markedly improves short- and long-term predictive accuracy, as shown in transfer-learning evaluations across diverse real-world networks (Yang et al., 2014).

Hybrid or composite centralities that integrate position (e.g., k-shell/coreness), local variation (e.g., degree differences among neighbors), or global spreading impact, such as entropy-weighted M-centrality, show superior performance in predicting influential spreaders and structural “weak points”—especially near spreading-threshold regimes or in networks with pronounced core-periphery structure (Ibnoulouafi et al., 2023).

---

In summary, degree centrality remains the foundational network centrality measure, uniquely combining computational efficiency, interpretability, and broad empirical relevance. Its extensions—generalized degree, m-order degree mass, cross-layer aggregations, and hybridizations with coreness measures—provide the necessary augmentations for addressing structural, multiplex, and dynamical complexities in modern network data (Csató, 2015, Bródka et al., 2012, Ibnoulouafi et al., 2023).