Profiler importance Ranking

[#1]

When running the XGBoost model the fit deatils of the model generate a feature importance ranking table. When running the profiler and subsequently the importance ranking gives another ranking of the predictor variables. Can somebody explain this different ranking of the features in the same ML model?

Which ranking do I have to choose in case I want to perform a feature selection, the model ranking or the profiler ranking?

Regards,

Lu

[#2]

When you say, "running the profiler and subsequently the importance ranking gives another ranking of the predictor variables," are you referring to using one of the Prediction Profiler commands shown here:

[#3]

Yes, I do

[#4 · Accepted solution]

XGBoost and the Prediction Profiler use completely different metrics for variable importance. The XGBoost add-in produces an index called Gain that is based on the splitting behavior of the underlying boosted trees. The Prediction Profiler uses resampling to produce an index of response variability against predictor variability. The two different rankings will not necessarily be the same.