GPs: Likelihoods for Zero-Inflated Count Data

1

Hi All,

Thanks for authoring a wonderful probabilistic programming language package. I was wondering if there any plans to add likelihoods for zero-inflated count data? E.g., ZIP (zero-inflated Poisson) and ZINB (zero-inflated negative binomial). Or is there a way to add custom likelihoods yourself?

Even if this means the means of inference becomes some sort of MCMC (e.g., HMC and its variants) it’d be a nice feature. PyMC3, btw, has ZIP support:

https://bwengals.github.io/gps-with-non-normal-likelihoods-in-pymc3.html

Many thanks ahead of time!

2

Hi, Pyro already has a zero-inflated Poisson distribution.

You should be able to implement additional zero-inflated distributions by forking the zero-inflated Poisson source code or by using pyro.distributions.MaskedMixture plus enumeration as in the observation distributions in the mixed-effect HMM example model.

1 Like
3

Thanks! Can this be used as a likelihood with a latent GP for the rate function and/or the probability of non-zero observations?

4

Yes, you should be able to fork the contrib.gp Poisson likelihood source code to use ZeroInflatedPoisson instead and use that new likelihood with the rest of the tools in pyro.contrib.gp.

1 Like