multi_predict() for glmnet to return mean count

[hfrick]

closes #63

library(poissonreg)
#> Loading required package: parsnip

data(seniors, package = "poissonreg")

spec <- poisson_reg(penalty = 0.1, mixture = 0.3) %>%
  set_engine("glmnet", nlambda = 15)
f_fit <- fit(spec, count ~ ., data = seniors)

# default: type = NULL -> "numeric"
our_pred <- predict(f_fit, seniors, penalty = 0.1)
our_pred
#> # A tibble: 8 × 1
#>   .pred
#>   <dbl>
#> 1 540. 
#> 2 740. 
#> 3 282. 
#> 4 387. 
#> 5  90.7
#> 6 124. 
#> 7  47.4
#> 8  65.0
identical(our_pred, predict(f_fit, seniors, penalty = 0.1, type = "numeric"))
#> [1] TRUE

# default: type = NULL
our_multi_pred <- multi_predict(f_fit, seniors, penalty = 0.1) %>% 
  tidyr::unnest(cols = .pred)
our_multi_pred
#> # A tibble: 8 × 2
#>   penalty .pred
#>     <dbl> <dbl>
#> 1     0.1 540. 
#> 2     0.1 740. 
#> 3     0.1 282. 
#> 4     0.1 387. 
#> 5     0.1  90.7
#> 6     0.1 124. 
#> 7     0.1  47.4
#> 8     0.1  65.0


seniors_x <- model.matrix(~ ., data = seniors[, -4])[, -1]
seniors_y <- seniors$count
glmnet_fit <- glmnet::glmnet(x = seniors_x, y = seniors_y, family = "poisson",
                          alpha = 0.3, nlambda = 15)
waldo::compare(
  our_pred$.pred,
  predict(glmnet_fit, seniors_x, s = 0.1, type = "response") %>% as.vector()
)
#> ✔ No differences

# This is now also the response aka mean count (instead of the linear predictor)
waldo::compare(
  our_multi_pred$.pred,
  predict(glmnet_fit, seniors_x, s = 0.1, type = "response") %>% as.vector()
)
#> ✔ No differences

^{Created on 2024-11-11 with reprex v2.1.0}

[hfrick]

CI failure on oldrel-4 (R 4.0.5) is due to a undeclared dependency of distributional on R 4.1 through a use of |>. That's been fixed in the dev version of distributional: mitchelloharawild/distributional@2a2db76

CI failure on devel is due to C errors when building MatrixStats but I am not sure how poissonreg depends on that or what to do about it. Hoping it will resolve itself. If not, it should be a separate PR.

> pak::pkg_deps_explain("poissonreg", "matrixStats", dependencies = TRUE)
                                                                          
x matrixStats

[simonpcouch]

🏄