The methods scale well (by ABC standards). Here is the output on real data with about 300 parameters to infer. Notice that the local parameters do vary quite a lot, and allowing for that variation allows for better inference of the global parameters.

We have found that both Over Sampling and Under Matching give significant improvements with mis-specified models or outliers, as these schemes make it easier to catch the tails of the posterior predictive.

Taken alone, this natural idea already gives us a gain of an order of magnitude.
(In this figure, lower is better.)

We exploit the model structure to accelerate the inference. We use a linear assignment algorithm to find the permutation of the synthetic data that best matches the observed data.

We consider ABC for hierarchical models, with both local and global parameters. Our motivating example is an SIR model of early Covid data across 94 French departements: some parameters are shared across all departements (global), others are specific to each departement (local).

We consider whether the communication systems of various species of tits, chickadees, and related birds, exhibit FD calls, a non-trivial composition of two base calls.

These are plotted on the known phylogeny: red nodes exhibit FD calls, grey nodes (or don't pass our rather stringent condition).

Enjoying discovering Singapore, and ready for BayesComp 2025!

I'll be giving a talk tomorrow (Wed) on Saddlepoint Monte Carlo and its application to exact Ecological Inference. Come say hi at 16:10 in room LT51 - it's a cool paper, I swear!

arxiv.org/abs/2410.18243

Sloppier-than-usual academic spam. 😬

I'd never thought of publishing a paper whose title would be my name. Sounds like a highly efficient plan, it would eliminate the whole headache of trying to pick a good title! 🙄

We also look at a neuroscience application, to analyse brain connectivity networks.

We applied this to find the underlying structure of the Infinito criminal network. It is surprisingly well resolved, and it recovers both known structure and interesting positions of certain members of the network.

The underlying tree might be a fully resolved phylogeny, or it might be a simple hierarchical structure, such as a small number of clusters.

We observe several adjacency matrices - say, multiple measurements, or networks defined by different features - with the same underlying tree.

Congratulations to Dr Charly Andral
who succesfully defended his PhD at Paris-Dauphine yesterday!
His doctoral work was supervised by Christian Robert and Randal Douc; I was delighted to be part of the jury.