#tidymodels — Public Fediverse posts

@jeremy-data.bsky.social biology based on #bioconductor (and I’m a BioJava contributor) - also model building in health informatics (I’d use #tidymodels and #vetiver now) vs Python

Preprint from Simon Wood on the new cross-validation smoothness estimation in #mgcv: https://arxiv.org/abs/2404.16490. It's a neat performant + data-efficient way to estimate GAMs based on complex CV splits (like spatial/temporal/phylo ones).

See ?NCV in latest {mgcv} for examples (https://cran.r-universe.dev/mgcv/doc/manual.html#NCV)

I might write a helper to convert {rsample}/{spatialsample} objects into mgcv's funny CV indexing structure.

#rstats #ml #tidymodels #mgcvchat @MikeMahoney218 @gavinsimpson @ericJpedersen @millerdl

#datatable #shiny #tidymodels #LatinR2023
@Posit @appsilon @RConsortium
200 attendees!!!

#RStats issues I'm struggling with that seem impossible to Google: Building a {brms} model within the {tidymodels} framework using {bayesian}.

The formula is inherently too complex (including splines and random effects) for the typical tidymodels workflow that involves recipes &c., so it must be added in at a later step. Two things:

1. Complex {brms} multivariate formulas seem to not be possible using {tidymodels}. E.g., literally multivariate or including phi after my formula via brms::bf(). It simply errors. :( This may just need some tweaking of {bayesian}'s scripts or waiting for an update since it's still fairly young.

2. Using {mgcv} random effect syntax like s(cat1, cat2, bs = "re") seems to not pick up as random effects in the model...I think? And I have never figured out if this is creating hierarchical random effects or not -- or if multilevel random effects just aren't possible in this syntax(?).

3. Using {lme4} random effect like (1 | cat1 / cat2) to ensure the hierarchy is preserved *does* retain random effects I can pull out of the model later using `ranef`, but for some absurd reason I cannot run this model through cross-validation or a myriad of other steps later because it seems to force-create a complex web of interacting factor levels that don't exist. E.g., if my random effects are '(1 | realm / biome)', this eventually fails because it'll look for tundra biome types in Africa for some absurd reason.*

Noticed this while trying to solve *separate* issues within broom.mixed:::tidy.brmsfit() -- that it seems to delete the names of all the fixed effects and return them as 'NULL' character strings (???), and its reliance on 'ranef' means it doesn't find the random effects using {mgcv} syntax.

That's my rambling mess of an essay for the day. Not sure how many of these are real issues or me simply not understanding how these packages differ or wot.

#brms #mgcv #tidymodels

* Almost wondering if this might even be a separate {tidymodels} issue right now. Every recipe no matter what seems to factor every single character column regardless of how the recipe is built. Hmmmm.