R: How to calculate and plot SHAP interaction values?

Question

It is pretty easy to compute and plot SHAP interaction values with Python: https://h1ros.github.io/posts/explain-the-interaction-values-by-shap/ or https://towardsdatascience.com/analysing-interactions-with-shap-8c4a2bc11c2a

I would like to do the same in R, preferably by using the packages kernelshap (compute SHAP values ) and shapviz (visualize SHAP). However, I struggle to find the equivalent for computing SHAP interaction values and how to plot them--can anyone please help?

Answer 1

Edit

The development version of "shapviz" now contains

• API for SHAP interactions (for "treeshap" and XGBoost)
• Dependence plots for main effects and pure interaction effects
• Still lacking: A good importance visualization

devtools::install_github("mayer79/shapviz")
library(shapviz)
library(xgboost)

dtrain <- xgb.DMatrix(data.matrix(iris[, -1]), label = iris[, 1])
fit <- xgb.train(data = dtrain, nrounds = 50)
x2 <- shapviz(fit, X_pred = dtrain, X = iris[, -1], interactions = TRUE)
sv_dependence(x2, "Petal.Length", interactions = TRUE)
sv_dependence(x2, "Petal.Length", color_var = "auto", interactions = TRUE)

enter image description here

Original answer

You can calculate SHAP values with packages "xgboost" and "treeshap" (and then plot them eg with "ggplot2".)

Example with XGBoost: predict(..., predinteraction=TRUE)

• Integration into "shapviz" is on the roadmap, and I will start to add such plots very soon): https://github.com/mayer79/shapviz/issues/6 (I will add your links)
• AFAIK, there is no theory yet of how to calculate Kernel SHAP interactions. Any reference is highly appreciated.