Mario Beraha

Very excited to give the next BNP webinar! I will be talking about some recent works on feature allocation models and several generalizations 🎉

📢I’m really proud of “Bayesian clustering of high-dimensional data via latent repulsive mixtures’’ just appeared on Biometrika, Advance articles, doi.org/10.1093/biomet…. Thank you to my terrific coauthors Lorenzo Ghilotti and @mberaha2.

📢The submission of contributed talks for the 14th International Conference on Bayesian Nonparametrics, UCLA (Los Angeles, US), June 23-27, 2025, is now OPEN! Deadline for submission: Dec 15, 2024.

Average temperatures by year, with the years listed in alphabetical order, and suddenly the climate crisis looks like a joke.

Warning: extremely difficult text to read below. Last night I surrendered my phone and signed a waiver. And then I sat with a small group and an Israeli military attaché and we watched the 45 minute compilation of Hamas videos from October 7th. 1/12

@matteogianella_, @bruno_guindani, @AlessandraGugl9, and I just released a new version of BayesMix, featuring an #Rstats interface! Extra features include full Windows support, more algorithms, more models, and more everything :) Try it out! github.com/bayesmix-dev/b…

As a bonus, we show that our model-based approach can be used to infer the cardinality of the dataset as well. This is another classical problem in computer science which was typically solved using a different data structure!

Hence, we adopt a pragmatic approach and propose the class "smoothed" estimators, which work very well in practice and are easy to compute! We extend the analysis to sketches obtained with multiple hash functions by drawing from the "multi-view" literature.

Together with Stefano Favaro and Matteo Sesia, we just arXiv'd our latest take on frequency and cardinality estimation from compressed (sketched) data: arxiv.org/abs/2309.15408

As a bonus, we show that our model-based approach can be used to infer the cardinality of the dataset as well. This is another classical problem in computer science which was typically solved using a different data structure!

Hence, we adopt a pragmatic approach and propose the class "smoothed" estimators, which work very well in practice and are easy to compute! We extend the analysis to sketches obtained with multiple hash functions by drawing from the "multi-view" literature.

@mberaha2 and I have just published the first project we started collaborating 5 years ago! Meanwhile we have taken onboard 4 more coauthors, and I thank them all. See Childhood obesity in Singapore: A Bayesian nonparametric approach, SMJ OnlineFirst

In particular, we show that, in some cases, this leads to a predictive distribution that depends on the sample size and the number of unique traits in the sample, similarly to what happens under the Pitman-Yor process prior!

Indeed, under a CRM prior, the predictive distribution of the number of "new" traits in an additional sample depends only on the sample size! We propose a new class of priors derived from the scaled subordinators by James et al. (2015).

Personal update. As of last week, I'm officially a Ph.D. in data science and computation. My thesis was on the statistical learning of RPMs, under the supervision of @AlessandraGugl9! We also arXived the first two papers from my postdoc, joint works with Stefano Favaro.