Neighborhoods are good communities

The communities of a social network are sets of vertices with more connections inside the set than outside. We theoretically demonstrate that two commonly observed properties of social networks, heavy-tailed degree distributions and large clustering coefficients, imply the existence of vertex neighborhoods (also known as egonets) that are themselves good communities. We evaluate these neighborhood communities on a range of graphs. What we find is that the neighborhood communities often exhibit conductance scores that are as good as the Fiedler cut. Also, the conductance of neighborhood communities shows similar behavior as the network community profile computed with a personalized PageRank community detection method. The latter requires sweeping over a great many starting vertices, which can be expensive. By using a small and easy-to-compute set of neighborhood communities as seeds for these PageRank communities, however, we find communities that precisely capture the behavior of the network community profile when seeded everywhere in the graph, and at a significant reduction in total work.

💡 Research Summary

The paper “Neighborhoods are good communities” investigates a surprisingly simple yet powerful observation about community structure in social and information networks: the immediate neighborhoods (egonets) of vertices often constitute high‑quality communities when the network exhibits two ubiquitous properties— a heavy‑tailed (power‑law) degree distribution and a large global clustering coefficient.

Theoretical contribution
The authors formalize the notion of conductance φ(S) as a measure of community quality and prove (Theorem 4.6) that in any undirected graph whose degree distribution follows a power law and whose global clustering coefficient κ is bounded away from zero, there must exist vertices whose egonets have conductance bounded by a function that decreases with the vertex degree. The intuition is that high κ guarantees many closed wedges (triangles) around a vertex, while a power‑law degree sequence guarantees the presence of high‑degree hubs. For a hub of degree d, the number of internal edges in its egonet scales as κ·d²/2, whereas the number of edges crossing the egonet boundary scales roughly as d·⟨k⟩. Consequently, for sufficiently large d the ratio of boundary edges to internal volume becomes small, yielding low conductance. This result extends the classic observation that a graph with κ = 1 is a disjoint union of cliques, showing that even moderate clustering together with heavy‑tailed degrees forces the existence of dense, low‑conductance local subgraphs.

Methodology
To validate the theory, the authors evaluate egonet communities on a diverse collection of real‑world networks (collaboration graphs, online social platforms, technological infrastructure, web link graphs) and on synthetic models (Chung‑Lu, Kronecker, stochastic block models). They compare egonet conductance against four baselines: (i) the Fiedler community derived from the second eigenvector of the normalized Laplacian (Cheeger bound), (ii) the best personalized PageRank (PPR) community obtained by sweeping over the PPR vector for many seed vertices, (iii) whisker communities (tiny dense subgraphs attached by a single edge), and (iv) METIS graph partitioning. Conductance, community size, and the network community profile (NCP) – the plot of minimum conductance versus community volume – are the primary evaluation metrics.

Empirical findings

1. Egonet quality – Across most datasets, egonet communities of size 10–200 vertices achieve conductance values (φ ≈ 0.1–0.2) comparable to or better than the Fiedler cut, confirming the theoretical prediction that local neighborhoods can be near‑optimal in the Cheeger sense.
2. NCP alignment – The conductance curve generated by the collection of egonet communities reproduces the characteristic “U‑shaped” NCP observed in prior work using exhaustive PPR sweeps. In particular, at small scales the egonet curve matches the lower envelope of the PPR curve, indicating that egonets capture the best bottlenecks that PPR would find with far more computation.
3. Seed efficiency for PPR – By selecting a modest set of low‑conductance egonets as seeds for the Andersen‑Chung‑Lang local expansion algorithm, the authors obtain PPR communities whose conductance profiles are virtually indistinguishable from those obtained by seeding every vertex. The total work (measured in push operations) drops to roughly 1–2 % of the exhaustive approach, demonstrating a dramatic speed‑up without sacrificing quality.
4. k‑core structure – Networks with high κ and power‑law degrees also possess large k‑cores (high‑minimum‑degree subgraphs). The authors show that egonet communities inside these cores retain low conductance, suggesting a hierarchical organization where dense cores contain many high‑quality local communities.

Limitations and discussion
The theoretical bound is a worst‑case guarantee; empirical conductance is often much lower, but the constants in the proof are loose, limiting precise quantitative predictions. Egonet‑based methods rely on the presence of high‑degree vertices; in graphs lacking hubs (e.g., trees or regular lattices) the approach yields weaker results. Moreover, conductance is only one aspect of community quality; the relationship of egonet communities to modularity, overlapping community detection, or semantic cohesion remains an open question.

Conclusion
The paper makes three key contributions: (1) it provides a rigorous proof that the combination of a power‑law degree distribution and a sizable clustering coefficient forces the existence of low‑conductance egonets, (2) it empirically demonstrates that these egonets are competitive with state‑of‑the‑art global methods on a wide range of networks, and (3) it shows that using egonets as seeds for local PPR expansion yields near‑optimal community profiles at a fraction of the computational cost. This work bridges a gap between theoretical graph properties and practical community detection, suggesting that in many large‑scale networks, simple local statistics are sufficient to uncover meaningful community structure without resorting to expensive global optimization.