Sensitivity of complex networks measurements

Complex networks obtained from the real-world networks are often characterized by incompleteness and noise, consequences of limited sampling as well as artifacts in the acquisition process. Because the characterization, analysis and modeling of complex systems underlain by complex networks are critically affected by the quality of the respective initial structures, it becomes imperative to devise methodologies for identifying and quantifying the effect of such sampling problems on the characterization of complex networks. Given that several measurements need to be applied in order to achieve a comprehensive characterization of complex networks, it is important to investigate the effect of incompleteness and noise on such quantifications. In this article we report such a study, involving 8 different measurements applied on 6 different complex networks models. We evaluate the sensitiveness of the measurements to perturbations in the topology of the network considering the relative entropy. Three particularly important types of progressive perturbations to the network are considered: edge suppression, addition and rewiring. The conclusions have important practical consequences including the fact that scale-free structures are more robust to perturbations. The measurements allowing the best balance of stability (smaller sensitivity to perturbations) and discriminability (separation between different network topologies) were also identified.

💡 Research Summary

The paper addresses a fundamental yet often overlooked problem in complex‑network research: the impact of incomplete and noisy data on the reliability of network measurements. Real‑world networks are rarely captured perfectly; limited sampling, measurement errors, and preprocessing artifacts introduce missing edges, spurious connections, and other distortions. Because virtually every analysis—characterization, modeling, prediction—relies on the underlying graph, it is essential to understand how such imperfections affect the metrics used to describe network structure.

To this end, the authors conduct a systematic sensitivity study involving six representative network models and eight widely used topological measurements. The models include Erdős‑Rényi random graphs, Watts‑Strogatz small‑world graphs, Barabási‑Albert scale‑free graphs, hierarchical modular graphs, spatially embedded graphs, and two empirical networks (a collaboration network and a protein‑interaction network). For each network they compute: average degree, average clustering coefficient, average shortest‑path length, global efficiency, betweenness centrality, assortativity, spectral radius, and modularity.

Three progressive perturbation schemes are applied to each original graph: (1) edge suppression – random deletion of a fraction p of existing edges; (2) edge addition – random insertion of new edges, preserving the same fraction p; and (3) edge rewiring – random removal of an edge followed by reconnection of its endpoints to two other nodes, thus preserving the total edge count while reshaping the topology. The perturbation intensity p is varied from 1 % to 30 % in steps, allowing a fine‑grained view of how metrics evolve under increasing distortion.

Sensitivity is quantified using the relative entropy (Kullback‑Leibler divergence) between the probability distributions of a given metric on the original network and on the perturbed version. A low KL value indicates that the metric is robust to the specific perturbation, whereas a high value signals strong sensitivity. In addition to robustness, the authors assess discriminability: the ability of a metric to separate different network models, measured by the average KL divergence between pairs of distinct models.

The results reveal two clear patterns. First, scale‑free networks exhibit the greatest resilience to edge suppression and rewiring. Because a few high‑degree hub nodes dominate the connectivity, random removal or relocation of edges rarely disrupts the core structure, leading to consistently low KL divergences across all three perturbation types. In contrast, random and small‑world networks are especially vulnerable to edge addition; the insertion of random shortcuts dramatically reduces average path length and alters clustering, producing large KL values. Second, among the eight metrics, the average clustering coefficient and average shortest‑path length emerge as the most stable across all perturbations while still maintaining high discriminability between models. Their global nature makes them less sensitive to local edge changes yet sufficiently informative to capture the characteristic differences of the underlying topologies.

Metrics such as betweenness centrality and spectral radius are highly sensitive: even modest edge modifications cause substantial shifts in their distributions, reflected in high KL values. Assortativity shows pronounced sensitivity to rewiring because the degree‑correlation pattern is easily disturbed when edges are reassigned. Global efficiency and modularity are moderately robust but have lower discriminative power compared to clustering and path length.

From a practical standpoint, the findings suggest concrete guidelines for researchers handling imperfect network data. When the data acquisition process is known to be noisy or incomplete, relying on clustering coefficient and average path length provides a more trustworthy characterization, reducing the risk of drawing erroneous conclusions from artefactual variations. Moreover, the demonstrated robustness of scale‑free structures implies that protecting hub nodes (e.g., through targeted monitoring or reinforcement) can be an effective strategy for preserving network functionality under attack or failure.

The authors conclude by emphasizing the importance of jointly considering robustness and discriminability when selecting network metrics. They also outline future research directions, including the study of targeted (non‑random) attacks, temporal networks where the topology evolves over time, and multivariate approaches that combine several metrics to build composite robustness profiles. Overall, the paper delivers a rigorous, quantitative framework for assessing measurement sensitivity, offering valuable insights for both theoretical investigations and applied network‑science tasks where data quality cannot be guaranteed.