Any way to speed up SHAP interaction values?

[Chrisjb]

When computing the interactions for use in dependence plots, we run:

shap_int <- predict(xgb_mod, (X_train), predinteraction = TRUE)

This takes a really, really long time when the xgboost model has lots of variables. Do you know of any way of speeding up the calculation? Perhaps by only letting it consider features with high average importance?

Thanks for the great work

[liuyanguu]

I suggest referring to discussions at https://github.com/slundberg/shap and the XGBoost package: https://github.com/dmlc/xgboost
After all we just get the interaction SHAP values from predict.xgb.Booster using predict(xgb_model, (X_train), predinteraction = TRUE)

[Chrisjb]

Thanks, will do. I ended up building a smaller XGBoost model based on the top n variables from the original - less variables reduces compute time exponentially