Koovely & Bovet's entropy method detects network phase changes in real time

Samuel Koovely and Alexandre Bovet have developed a novel method for detecting structural phase changes in temporal networks using information theory. Their paper, published on arXiv, extends the conditional entropy of heat diffusion—previously only defined for static graphs—to dynamic temporal networks. The key theoretical contribution is proving that this entropy measure is monotonic over time, creating an information-theoretic analog of the second law of thermodynamics for inhomogeneous diffusion. This allows researchers to quantify disorder in evolving systems like social contact networks or communication graphs.

The authors also introduce a local version of the conditional entropy that works over finite time windows, providing a sensitive signal for change-point detection. They validated the methodology on synthetic benchmarks, showing superior performance compared to existing nonparametric snapshot-based methods. A real-world test on a temporal contact network from a French primary school demonstrated practical utility. Furthermore, they show that using detected change points to segment the network and then apply community detection per interval yields higher-quality, more interpretable clustering results.