Bayesian nonparametric clustering as a community detection problem

摘要: Abstract A wide class of Bayesian nonparametric priors leads to the representation distribution observable variables as a mixture density with an infinite number components. Such induces clustering structure in data. However, due label switching, cluster identification is not straightforward posteriori and some post-processing MCMC output usually required. Alternatively, observations can be mapped on weighted undirected graph, where each node represents sample item edge weights are given by posterior pairwise similarities. It shown how, after building particular random walk such it possible apply community detection algorithm, known map equation, leading minimisation expected description length partition. relevant feature this method that allows for quantification uncertainty classification.