Inconsistent predictions

[Hayakawa94]

Hi

I need some help with the R xgboost model. I have built a claim severity model using the reg:gamma objective. When assessing the predictions, I noticed different predictions being outputted when iterationrange = c(1,1) is specified. The result is below:

(predict(xgb_model,
newdata = as.matrix(gbm.data%>% select( xgb_model$feature_names ) ) ,
iterationrange = c(1,1)
) / predict(xgb_model,
newdata = as.matrix(gbm.data%>% select( xgb_model$feature_names ) )
) ) %>% summary

Min. 1st Qu. Median Mean 3rd Qu. Max.
0.6827 0.9903 0.9990 0.9987 1.0072 1.4092

Which is the correct prediction and which method is used to compute SHAP?

Thanks in advance

[Hayakawa94]

It appears that using the predict function without specifying iterationrange = c(1,1) includes trees from the early stopping rounds in the predictions. For example, if nround = 100, early stopping rounds = 10, and 80 trees were built, the predict function would use 90 trees instead of the 80 that were actually built. Could someone clarify if this would impact the SHAP computations?

[david-cortes]

If you are using the latest development version of XGBoost, or if you installed it from GitHub, note that the interpretation of iterationrange = c(1,1) has changed, and the docs have been updated to reflect the new behavior:

xgboost/R-package/R/xgb.Booster.R

Line 118 in 2d1c26b

#' @param iterationrange Sequence of rounds/iterations from the model to use for prediction, specified by passing