how to get Shap interactions for LightGBM?

[pecto2020]

Your package is great, and very easy to use within tidymodels framework. I was wondering if it is possible to calculate interactions for LightGBM. I would like to use that instead of the heuristic (which is an amazing solution tho) in sv_dependence. I've seen that for Xgboost is possible and there is a param Interaction = T to set in shapviz.base. Any solution workaround for LightGBM?

[mayer79]

Unfortunately not via TreeSHAP in LightGBM. But you could crunch interactions via the {treeshap} package.

[mayer79]

microsoft/LightGBM#6058 @hbaniecki :-)

[hbaniecki]

I assume it involves hacking C++ code, which I can't help with :/

[mayer79]

Oh, hmm...

[pecto2020]

Tried to use treeshap but got an error #> Error in S_inter[, v, color_var]: subscript out of bounds

Here's the code

library(tidymodels)
#> Warning: package 'tidymodels' was built under R version 4.2.3
#> Warning: package 'broom' was built under R version 4.2.3
#> Warning: package 'dials' was built under R version 4.2.3
#> Warning: package 'dplyr' was built under R version 4.2.3
#> Warning: package 'ggplot2' was built under R version 4.2.3
#> Warning: package 'parsnip' was built under R version 4.2.3
#> Warning: package 'recipes' was built under R version 4.2.3
#> Warning: package 'tibble' was built under R version 4.2.3
#> Warning: package 'tidyr' was built under R version 4.2.3
#> Warning: package 'tune' was built under R version 4.2.3
#> Warning: package 'workflowsets' was built under R version 4.2.3
#> Warning: package 'yardstick' was built under R version 4.2.3
library(shapviz)
#> Warning: package 'shapviz' was built under R version 4.2.3
library(treeshap)
library(lightgbm)
#> Warning: package 'lightgbm' was built under R version 4.2.3
#> Loading required package: R6
#> 
#> Attaching package: 'lightgbm'
#> The following object is masked from 'package:dplyr':
#> 
#>     slice
library(datasets)
library(bonsai)
#> Warning: package 'bonsai' was built under R version 4.2.3

# Use the fifa20 dataset 
fifa20 <- fifa20$data %>% 
            select(-work_rate) %>%
            bind_cols(data.frame(target = fifa20$target)) 
# Split the data
set.seed(123)
split <- initial_split(fifa20)
train <- training(split)
test <- testing(split)

# Recipe 
rec <- recipe(target ~ ., data = train)

# Model specification
boost_spec <- boost_tree(
  mode = "regression", 
  trees = 200, 
  tree_depth = 6
) %>%
  set_engine("lightgbm") %>%
  set_mode("regression")

# Workflow
workflow <- workflow() %>% 
  add_recipe(rec) %>%
  add_model(boost_spec)

# Fit the model
boost_model <- workflow %>% fit(data = train)

# Create shap object with shapviz
shap_lgbm <- shapviz(extract_fit_engine(boost_model), 
                    as.matrix(test %>% select(-target)),
                    test %>% select(-target))

# Create unified model representation
unified_lgbm <- treeshap::lightgbm.unify(extract_fit_engine(boost_model), train %>% select(-target))

# Derive interactions
interactions_lgbm <- treeshap::treeshap(unified_lgbm, test %>% select(-target), interactions = T, verbose = 0)

# Plot dependences
shap_lgbm$S_inter <- interactions_lgbm$interactions

sv_dependence(shap_lgbm, v = "overall", interactions = T,  color_var = "height_cm")
#> Error in S_inter[, v, color_var]: subscript out of bounds

dim(shap_lgbm$S_inter)
#> [1]   54   54 4570

[mayer79]

An interaction cannot be assigned to a shapviz object, so this code here is wrong:

shap_lgbm$S_inter <- interactions_lgbm$interactions

This works, but I would decompose less rows and divide the response by 1e6 (or so):

shap_lgbm <- shapviz(interactions_lgbm)
top4 <- names(head(sv_importance(shap_lgbm, kind = "no"), 4))
sv_interaction(shap_lgbm[1:1000, top4])
sv_dependence(shap_lgbm, v = "overall", color_var = top4, interactions = TRUE)