I was wondering on some more information on how gbm handles missing data following this tutorial.
Decision trees are robust to outliers and handle missing data very well using surrogate splits. The latter feature is a very useful one. Suppose we train a decision tree to predict variable Y using features A, B, C. If we receive new data that only had valid values for A,B, we would be in trouble and might not be able to predict Y. Tree models are trained to include surrogate splits, so that if a variable C is missing in a new data point, the decision defers to another variable C' that is highly correlated with C and proceeds with the prediction
I collected crop yield data across many locations which are identified by IDs.
I am trying to run
gbm package with learning rate of
tc = 2 and using climate variables and the location IDs as predictors.
If I want to run the fitted model to a a new location with a different ID which was not present in the model, how would the boosted regression trees predict for this new ID?