Bayesian Clustering : Hierarchical, contiguity constrained (Etienne Côme, Cosys, Univ. Gustave Eiffel)

Standard clustering methods often fail to account for the topological structure of spatial or network data, leading to fragmented results. This talk presents a Bayesian Hierarchical Clustering framework that strictly enforces contiguity constraints, ensuring that every cluster forms a connected subgraph. We will introduce a generative model based on spanning trees that allows exact posterior probability computation of partitions. We will also cover the extraction of Bayesian dendrograms, which provide a probabilistic take on classical hierarchical outputs. These approaches will be demonstrated using the greed and gtclust R packages on both synthetic and real-world datasets in both the contiguity-constrained and unconstrained settings.