A0472
Title: Bayesian multi-resolution clustering via infinite latent factors
Authors: Lorenzo Schiavon - Ca Foscari University of Venice (Italy) [presenting]
Mattia Stival - Ca Foscari University of Venice (Italy)
Abstract: In many scientific fields, a critical task is to cluster subjects based on a potentially vast set of features. A fundamental challenge in model-based clustering is the trade-off between the resolution of the inferred clusters and the parsimony of the model. Current nonparametric approaches often require a pre-specified resolution level, demanding extensive parameterization to capture fine-grained structures and offering no mechanism to explore cluster hierarchies. To overcome these limitations, a novel multi-resolution clustering approach is introduced using an infinite mixture model with kernels organized in a multiscale framework. The method, through a careful specification of mixture weights, naturally incorporates exogenous information to guide the formation of cluster hierarchies while maintaining flexibility. The theoretical properties of the model are investigated, and an elegant and parsimonious formulation is proposed based on an infinite factorization, which allows for efficient posterior inference via a Gibbs sampler. The practical advantages of the approach are shown on synthetic data and through challenging real-world applications, revealing multi-level grouping patterns in survey responses and gene expression data.
