Column contribution equivalent for XGBoostHi @russ_wolfinger ,
I have a question regarding the interpretation of the information contained in the Actual by Predicted plots part of the XGBoost report. Within it, is an "Importances" sub table that contains the Gain and Cover statistics for the model of interest.
What I'm interested in, is to better understand which of these corresponds to the column contributions information that's given in platforms like boosted tree and bootstrap forest. To me, it seems like the Cover statistic is the best analogue to the column contributions, but maybe I'm mistaken.
As an example, below is an output I have for a model I ran. The splits column is clear, and the Gain is also well explained by the help PDF, but the Cover column, it seems if scaled correctly would be the equivalent of column contributions. I made a column called Contribution? that is the cover divided by the Col Sum(Cover). Am I interpreting the results correctly?
Thanks!,
DS
Re: Column contribution equivalent for XGBoostHi @russ_wolfinger ,
Thanks for the feedback, I will certainly review the data that way and compare the different results. A direct comparison might not be so trivial, but it seems that it's still somehow conceptually related, and I just need to find the best way to evaluate that given my specific situation.
Thanks!,
DS
Re: Column contribution equivalent for XGBoostHi @DS,
The cover statistic is more related to the number of observations that flow into nodes created by splitting on the variable. Suggest web searching "xgboost cover" to see several informative explanations online.
I believe Column Contributions from other platforms is more directly related to sums of squares and there may not be an easy transform to the XGBoost output. Standardizing Cover like you have done certainly seems reasonable in any case.
What I typically do is right click > Save as Data Table, log transform all three stats, assign the variable name as a label, then do multivariate to plot them against each other.
