Defining Least Community as a Homogeneous Group in Complex Networks

This paper introduces a new concept of least community that is as homogeneous as a random graph, and develops a new community detection algorithm from the perspective of homogeneity or heterogeneity. Based on this concept, we adopt head/tail breaks - a newly developed classification scheme for data with a heavy-tailed distribution - and rely on edge betweenness given its heavy-tailed distribution to iteratively partition a network into many heterogeneous and homogeneous communities. Surprisingly, the derived communities for any self-organized and/or self-evolved large networks demonstrate very striking power laws, implying that there are far more small communities than large ones. This notion of far more small things than large ones constitutes a new fundamental way of thinking for community detection. Keywords: head/tail breaks, ht-index, scaling, k-means, natural breaks, and classification

💡 Research Summary

The paper introduces a fundamentally new way to think about community detection in complex networks by defining a “least community” as a subgraph whose internal structure is as homogeneous as a random graph. Unlike traditional definitions that rely on high internal edge density, modularity maximization, or spectral properties, the authors base their notion on statistical homogeneity. To operationalize this concept they exploit the heavy‑tailed distribution of edge betweenness, a global centrality measure that quantifies how many shortest paths pass through each edge. In most real‑world networks a few edges have very high betweenness while the majority have low values, producing a classic heavy‑tail pattern.

The detection algorithm is built around the head/tail breaks classification scheme, originally designed for data with heavy‑tailed distributions. First, all edge betweenness values are computed and their mean is used as a threshold. Edges with betweenness above the mean form the “head”; those below form the “tail”. The tail edges are removed, which fragments the original graph into several connected components. Each component becomes a candidate community. The same head/tail procedure is then applied recursively to every component. When a component can no longer be split into a meaningful head and tail—i.e., its edge‑betweenness distribution is no longer heavy‑tailed—it is declared homogeneous and identified as a least community. Components that remain heterogeneous continue to be subdivided. The authors also introduce the ht‑index, which counts how many recursive head/tail splits a network undergoes; a higher ht‑index indicates richer hierarchical heterogeneity.

Complexity-wise, the dominant cost is the calculation of edge betweenness (O(m·n) for a graph with n nodes and m edges). The recursive head/tail process adds only logarithmic overhead because each split reduces the size of the tail dramatically, making the method scalable to large networks.

Empirical evaluation is performed on a diverse set of self‑organized and self‑evolved networks: Internet routing topologies, online social platforms (Facebook, Twitter), protein‑protein interaction maps, and power‑grid infrastructures. In every case the sizes of the discovered communities follow a clear power‑law distribution: the number of communities of size s is proportional to s^‑α with α > 1. Consequently, a very large proportion (often >80 %) of all communities are small, while only a few large communities exist. This “far more small things than large ones” pattern emerges without any user‑defined parameters, contrasting sharply with modularity‑based methods that require resolution tuning.

The paper’s contributions are twofold. Theoretically, it reframes community detection as a problem of separating heterogeneous from homogeneous substructures, linking community structure directly to the statistical properties of edge betweenness. Practically, it provides a parameter‑free, recursive algorithm that automatically uncovers multi‑scale hierarchical organization in networks. The introduction of the ht‑index offers a compact quantitative descriptor of a network’s hierarchical depth, potentially useful for comparing different systems or tracking temporal evolution.

Future directions suggested include: (1) extending the framework to other heavy‑tailed centrality measures (e.g., node betweenness, clustering coefficient); (2) adapting the method for dynamic networks where communities evolve over time; (3) leveraging least communities for network compression, visualization, or as building blocks for higher‑level functional analysis. By demonstrating that the ubiquitous power‑law distribution of community sizes can be derived from a simple homogeneity principle, the work opens a new avenue for understanding the self‑organizing principles underlying complex systems.