A Bayesian approach to ensemble clustering
Federico Maria Quetti, Silvia Figini, Elena Ballante · Statistics · 2025
The paper presents a novel Bayesian approach to clustering techniques. A new method is proposed in an ensemble scheme for clustering, aimed at improving results in terms of robustness and interpretability. Our approach is organized in two steps. Firstly, a clustering algorithm is introduced by deploying a bagging scheme with the proper Bayesian bootstrap resampling method. The algorithm is based on K-means for elicitation of the prior for proper Bayesian bootstrap deployment. The clustering results for bootstrap replicas are finally aggregated as soft assignments. Subsequently, a novel method to determine the optimal number of clusters is derived, by introducing measures of uncertainty on assignments based on Shannon entropy. Empirical results are provided on simulated and real data showing the potential advances for applications, obtained in terms of clustering quality, stability, and correct choice of the number of clusters.