As in my recent experience I didn't found a full tutorial, walkthrough or example of how to perform a step by step "personalized" XGBoost in R, I decided to upload this. Hope the code will be useful to someone.
What I mean by "personalized"? I mean that you could use any loss-function (evaluation metric, gain function...) to perform a fully parallelized Hyperparameter Tuning and then use XGBoost with those hyperparameters for whatever you want: prediction, feature selection, support for descriptive or causality analysis and so on. In xgboost_tutorial_dv.R I leave a tutorial to:
- Hyperparameter Tuning XGBoost using any evaluation metric you want with parallelized computation.
- With some recommendations about the Hyperparameter Space
- Optimal iteration for XGBoost (partially parallelized).
- Personalized Cross-Validation to be sure and to explore more about predicting power of the model (fully parallelized).
- Feature importance computation from where you can easily perform Feature Selection.
- Computation of an "interpretable model" from our XGBoost (thanks to
AppliedDataSciencePartners/xgboostExplainer)
Also I uploaded Health Survey Ecuador - 2018 - as presented by publisher.rar, databases you could use to reproduce my results if you want. Those are bases from the National Health Survey from Ecuador, survey conducted by the National Statistical Office.