We present a probabilistic inference framework based on mean field and belief propagation techniques for estimating risk of infection from proximity tracking data used in contact tracing. The algorithm, using the SIR inference model, is described in this short paper, where you can find more information about our approach.
Demonstrations of the package and of the inference procedure can be found on these google colab notebooks:
-
A set of demos and sanity checks: Scatterplot compares the probabilities estimated by mean field and DMP with the frequencies obtained by Monte Carlo runs, showing the efficiency of the algorithm in sampling.
-
Demo 1 and Demo 2 are basic comparisons of the efficiency between choosing subjects to test via contact tracing, random sampling and with the inference approach.
-
Demo 3 compares the efficiency of quarantine strategies using contact tracing as well as mean-field and message-passing inference in simple SIR models. It was used to generate the image shown above.
-
Demo 4 is another such demonstration ran on large realistic networks generated by the agent-based model developed by the Pathogen Dynamics group in the Big Data Institute of Oxford.
- Lenka Zdeborova presented this work online at the "ELLIS against Covid-19" event on May 6, 2020, see slides or the talk on youtube
Antoine Baker, Florent Krzakala, Marc Mézard, Stefano Sarao Mannelli, Maria Refinetti and Lenka Zdeborová