Reminder: All About that Bayes seminar series - Bayesian Mixture Models, February 13, 2026

[All About That Bayes]

Dear colleagues,

Quick reminder for the seminar “All About that Bayes” that takes place on 13 February 2026! Below you can find the full program.

The seminar is open to everyone, but please confirm your attendance by registering via   this link.

Looking forward to seeing many of you there!

Best regards,
Isabelle Albert & Kaniav Kamary
On behalf of the Specialized Group “Bayesian Statistics” of the French Statistical Society (SFdS)
-----------------------------------------------------------------------------------------------------
Theme: Bayesian Mixture Model
Date: Friday, 13 February 2026
Time: 2:00 – 5:00 PM
Location: Y. Cauchois Room, Institut Henri Poincaré (IHP)- Sorbonne Université / CNRS, Paris

Speakers:
Darren Wraith (School of Public Health and Social Work, Center for Data Science, Queensland Institute of Technology)
Title: Efficient adaptive importance sampling in challenging geometric and high dimensional settings
Abstract: A number of adaptive importance sampling approaches are based on proposals
 using Gaussian or Student-t mixture distributions which adapt to a target density. In a Bayesian
 setting and for problems in cosmology, computational biology, or climate modeling-the
computation of the likelihood can be expensive. Standard mixture based approaches may
 generate redundant samples when mixture components overlap, repeatedly evaluating the
 likelihood in the same posterior regions without improving the estimate. We examine different
 approaches which can be used to reduce likelihood evaluations while maintaining unbiasedness
 and low variance. One approach is to use a point-level thinning mechanism that assigns
 retention probabilities to drawn samples based on responsibility calculations before the likelihood
 evaluation. Points falling in densely covered regions are stochastically discarded, with importance
 weights corrected by the inverse retention probability to preserve unbiasedness. We also explore
 approaches which can force components to be separate from each other using a repulsive penalty
 applied to the objective function. Both methods preserve the key advantages of mixture based
 adaptive approaches while dramatically reducing computational cost.  We discuss theoretical
 properties including unbiasedness under thinning, convergence of the mixture adaptation,
computational complexity, and practical implementation strategies including variance reduction
 techniques. The methods are applicable to any setting where likelihood or posterior evaluations
 dominate computational cost.  We illustrate the performance of the approaches in challenging
 geometric and high dimensional settings.

Jackie Rao (MRC Biostatistics Unit, University of Cambridge)
Title: Federated Variational Inference for Bayesian Mixture Models
Abstract: We present a one-shot, unsupervised federated learning approach for Bayesian
model-based clustering of large-scale binary and categorical datasets, motivated by the need to
identify patient clusters in privacy-sensitive electronic health record (EHR) data. We introduce
a principled 'divide-and-conquer' inference procedure using variational inference with local
merge and delete moves within batches of the data in parallel, followed by 'global' merge moves
across batches to find global clustering structures. We show that these merge moves require
only summaries of the data in each batch, enabling federated learning across local nodes
without requiring the full dataset to be shared, thus preserving privacy. Empirical results
on simulated and benchmark datasets demonstrate that our method performs well relative to
comparator clustering algorithms, achieving significant computational speedups.
We validate the practical utility of the method by applying it to a large-scale British primary
care EHR dataset comprising 289,821 individuals to identify clusters of individuals with
common patterns of co-occurring conditions (multimorbidity).

Martin Metodiev (Laboratoire de Mathématiques Blaise Pascal, Université Clermont Auvergne)
Title: Easily Computed Marginal Likelihoods for Multivariate Mixture Models Using the THAMES Estimator
 Abstract: We present a new version of the truncated harmonic mean estimator (THAMES) for
univariate or multivariate mixture models. The estimator computes the marginal likelihood from
 Markov chain Monte Carlo (MCMC) samples, is consistent, asymptotically normal and of finite variance.
In addition, it is invariant to label switching, does not require posterior samples from hidden allocation
 vectors, and is easily approximated, even for an arbitrarily high number of components. Its computational
 efficiency is based on an asymptotically optimal ordering of the parameter space, which can in turn
be used to provide useful visualisations. We test it in simulation settings where the true marginal
likelihood is available analytically. It performs well against state-of-the-art competitors, even in
multivariate settings with a high number of components. We demonstrate its utility for inference
 and model selection on univariate and multivariate data sets.