Hierarchical community detection via maximum entropy partitions and the renormalization group

Abstract: Identifying meaningful structure across multiple scales remains a central challenge in network science. We introduce Hierarchical Clustering Entropy (HCE), a general and model-agnostic framework for detecting informative levels in hierarchical community structures. Unlike existing approaches, HCE operates directly on dendrograms without relying on edge-level statistics. It selects resolution levels that maximize a principled trade-off between the entropy of the community size distribution and the number of communities, corresponding to scales of high structural heterogeneity. This criterion applies to dendrograms produced by a wide range of clustering algorithms and distance metrics, including modularity-based and correlation-based methods. We evaluate HCE on synthetic benchmarks with varying degrees of hierarchy, size imbalance, and noise, including LFR and both symmetric and asymmetric multiscale models, and show that it consistently identifies partitions closely aligned with ground truth. Applied to real-world networks in social and neuroscience systems, HCE reveals interpretable modular hierarchies that align with known structural and functional organizations. As a scalable and principled method, HCE offers a general, domain-independent approach to hierarchical community detection with potential applications across biological, social, and technological systems.