Robustness of Social Networks: Comparative Results Based on Distance Distributions

Given a social network, which of its nodes have a stronger impact in determining its structure? More formally: which node-removal order has the greatest impact on the network structure? We approach this well-known problem for the first time in a setting that combines both web graphs and social networks, using datasets that are orders of magnitude larger than those appearing in the previous literature, thanks to some recently developed algorithms and software tools that make it possible to approximate accurately the number of reachable pairs and the distribution of distances in a graph. Our experiments highlight deep differences in the structure of social networks and web graphs, show significant limitations of previous experimental results, and at the same time reveal clustering by label propagation as a new and very effective way of locating nodes that are important from a structural viewpoint.

💡 Research Summary

The paper tackles the classic problem of identifying the most structurally important nodes in large networks by measuring how node removal affects the network’s distance distribution and the number of reachable node pairs. Traditional robustness studies have relied on simple metrics such as graph diameter or the size of the giant component, but these are either computationally infeasible for massive graphs or insufficient to capture subtle structural changes. Leveraging the recently introduced HyperANF algorithm, the authors are able to approximate the neighbourhood function N_G(t) and the cumulative distance distribution H_G(t) for graphs containing billions of nodes and hundreds of billions of edges with provable error bounds.

A removal order (denoted ≺) is defined, and nodes are deleted sequentially according to this order until a given fraction θ of arcs has been removed. The resulting graph G(≺, θ) is compared with the original graph G by examining (1) the ratio of still‑reachable pairs to the original number of reachable pairs, and (2) the change in the distance distribution. Several divergence measures are considered—relative change in average distance, relative change in harmonic diameter, Kullback‑Leibler divergence, and L₁/L₂ norms—but the authors settle on the relative change in average distance δ(P,Q)=μ_Q/μ_P−1 because it is intuitive and yields consistent rankings across all experiments.

Four node‑removal strategies are evaluated:

1. Random – a baseline that selects nodes uniformly at random.
2. Largest‑degree first – removes nodes in decreasing out‑degree order, representing a simple degree‑centrality approach.
3. PageRank – removes nodes according to their PageRank scores, a global centrality measure based on a Markov chain.
4. Label‑propagation clustering – first clusters the (symmetrised) graph using the label‑propagation algorithm, then iteratively removes, from each cluster in decreasing size order, the node with the highest number of edges to other clusters. This strategy targets “bridge” nodes that connect densely‑connected communities.

The experimental suite comprises a mixture of web graphs (snapshots of the .it, .uk, .com domains, among others) and social networks (the Hollywood co‑appearance graph, LiveJournal friendship graph, Flickr interaction graph, DBLP co‑authorship network, etc.). Web graphs typically display a classic small‑world pattern with a few extremely high‑degree hubs and relatively low clustering coefficients, while social graphs exhibit higher clustering, more uniform degree distributions, and many alternative short paths.

Key findings:

• Random removal has negligible impact on both distance distribution and reachable‑pair ratio for all datasets, confirming that large networks are robust to indiscriminate loss.
• Largest‑degree and PageRank removals dramatically increase average distance and reduce reachable pairs in web graphs; even a 1 % removal of arcs (θ≈0.01) can raise the average distance by >20 % and cut the reachable‑pair ratio by roughly 10 %. In contrast, the same strategies cause only modest changes (<5 % in average distance) in social graphs.
• The label‑propagation clustering strategy is the most disruptive for web graphs. Removing just 2 % of arcs (θ≈0.02) via this method can increase average distance by >30 % and halve the reachable‑pair ratio. This effect stems from the elimination of inter‑cluster “bridge” nodes that hold the small‑world shortcuts together. For social networks, however, the clustering‑based removal produces almost no measurable change in the distance distribution, underscoring their inherent resilience.

These results reveal a fundamental structural divergence between web graphs and social networks that is not captured by standard scale‑free models, which treat both as having similar degree‑based properties. Web graphs rely heavily on a handful of hub nodes; their removal quickly fragments the network and inflates path lengths. Social networks, by virtue of high clustering and abundant alternative routes, maintain short average distances even after targeted attacks on high‑centrality nodes.

The authors argue that distance‑distribution‑based robustness analysis offers a richer picture of network health than diameter or component‑size metrics alone, because it reflects the average accessibility of all node pairs. Practically, the findings suggest that protecting high‑PageRank or high‑degree hubs is crucial for maintaining the performance of web‑scale services (search, crawling, content delivery), whereas social platforms should focus on safeguarding community‑level connectivity rather than individual “influencers.”

In summary, the paper introduces a scalable methodology for approximating distance distributions on massive graphs, systematically evaluates several node‑removal heuristics, and demonstrates that web graphs and social networks exhibit markedly different robustness profiles. The work challenges the adequacy of simple scale‑free assumptions and highlights the importance of considering full distance‑distribution dynamics when assessing network resilience.