Bayesian nonparametric models such as Cox processes are a standard way to model event-based data, but their intensity functions are usually estimated with expensive MCMC sampling or approximate inference. In a new arXiv preprint, Joshua Rooijakkers and Edwin Fong argue that these methods often place priors over indirect latent structures, which makes them hard to interpret and costly to run.
Instead, the Martingale Posterior Cox Process (MPCP) treats the unobserved events themselves as the quantity to infer, given the observed event history. This removes the need for complex priors while keeping the flexibility of nonparametric modeling. The authors pair this perspective with a copula-based predictive update designed to be GPU-friendly and scalable, and they provide theoretical guarantees for the convergence of the predictive resampling algorithm.
In simulations and real data, MPCP reports excellent performance while requiring only a fraction of the computational cost of competing Bayesian approaches. Since this digest draws on a single preprint, those results have not yet been independently verified; the claims are the authors' own.