Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

BG/NBD Model in PyMC

See original GitHub issue

Hey! I have recently drafted a BG/NBD model in PyMC, see https://juanitorduz.github.io/bg_nbd_pymc/ In particular I have also added the possibility to add time-invariant regressors (even though the sampler and/or model parametrisation need some improvements). Other models from the lifetimes package can be ported to PyMC in a similar way (because the lifetimes package has all the log-likelihood functions written in numpy already).

Issue Analytics

State:
Created a year ago
Comments:6 (4 by maintainers)

Top GitHub Comments

1reaction

juanitorduzcommented, Apr 17, 2022

That sounds very exiting! Let me know if you need som input or want me to test some of the ongoing features. As a side remark, I also did a small re-write in pymc for the gamma-gamma model, see https://juanitorduz.github.io/gamma_gamma_pymc/

1reaction

ColtAllencommented, Mar 22, 2022

Hey @juanitorduz ,

This is great! I really hope you can join the Zoom kickoff call for this project next Sunday. Information to join is in a Discussion post in this repo as well as here:

https://github.com/CamDavidsonPilon/lifetimes/issues/414#issuecomment-1073247582

I’ve been working on a BG/NBD model as well, but will be experimenting with ADVI or SVGD for inference because MCMC training can go on for days, compared to minutes with ADVI. However, MCMC is also the only way to derive the true posterior distribution (rather than an approximation as is the case with ADVI or SVGD). A more performant option for MCMC may be Stan, which prophet uses for inference of its time series model parameters. @alexpavlakis wrote a script in Stan for the Gamma-Gamma model, but I haven’t tested it yet in CmndStanPy.

As you mentioned, the tricky thing about these algorithms is choosing an appropriate prior, and my goal is to make this library as user-friendly as possible. Fortunately in the case of the Pareto/NBD model, there is a research paper on this very topic:

Worth the effort? Comparison of different MCMC algorithms for estimating the Pareto/NBD model

I’ve saved several research papers on covariates and other modeling alternatives and will be posting links to the README soon. None of them are pay-walled, but to be safe I can also look into uploading my saved copies to the repo. One of these papers even covers a Pareto/NBD model with time-varying covariates, but it is immensely more compute-intensive than the standard Pareto/NBD and only performs significantly better for customers with lifetimes of 3 years or more.