Support for LightGBM multiclassova Multi-Class (one-vs-all) Objective

[JavOrraca]

I recently discovered that LightGBM has a multiclassova objective function (multi-class one-vs-all) that treats multi-class classification as a set of binary classification tasks. Getting quite decent results with this approach when passed directly to lightgbm::lgb.train(), but I'm not sure if the internal process_objective_function() overrides any objective function manually inserted in set_engine(), for example:

set_engine("lightgbm", params = list(objective = "multiclassova"))

Thanks for your help!

[simonpcouch]

Thanks for the issue! You can set lightgbm params by passing the desired parameter directly to set_engine() rather than as an element of set_engine(param):

library(bonsai)
#> Loading required package: parsnip

mtcars$gear <- as.factor(mtcars$gear)

spec <-
  boost_tree(mode = "classification", engine = "lightgbm") %>%
  set_engine("lightgbm", objective = "multiclassova", num_classes = length(unique(mtcars$gear)))

fit(spec, gear ~ ., mtcars)
#> parsnip model object
#> 
#> LightGBM Model (1 tree)
#> Objective: multiclassova (3 classes)
#> Fitted to dataset with 10 columns

^{Created on 2024-11-07 with reprex v2.1.1}

Will leave this issue open as this surfaced a few issues.

1. I thought this was more clearly documented but I didn't find the relevant docs in the first 2 pages I looked to
2. I thought train_lightgbm() handled an explicit params argument more gracefully than it seems to.
3. process_objective_function() could set num_classes for the user even with an explicitly-set objective function.

[JavOrraca]

Ah, thank you!