Xgboost Feature Importance

xgb_feature_importance.R

# xgboost feature importance ----------------------------------------------------------------

    #' Author: Taylor Van Anne
    #' 7/25/2017
    #' 
    #' This script is a simple demonstration of:
    #' 1) using cross validation to determine the optimal number of iterations for your xgboost model
    #' 2) runs xgboost with that number of iterations
    #' 3) extract the feature importance from the resulting model



    library(xgboost)
    set.seed(4)
    
    
    # set up data
    my_iris <- iris    
    just_features <- my_iris[, setdiff(names(my_iris), "Species")]
    just_labels <- as.integer(as.factor(my_iris$Species)) - 1
    
    
    # build a DMatrix
    my_iris_dmat <- xgboost::xgb.DMatrix(as.matrix(just_features), label=just_labels)
    
    
    
    # general params for both CV and real modeling -- change your objective function to match your situation
    myxgb_params <- list(objective = "multi:softmax", num_class=3, nthread=1)
    
    
            # list of additional parameters:
            # http://xgboost.readthedocs.io/en/latest/parameter.html
    
            # bonus material - difference between 'objective' and 'feval':
            # https://stackoverflow.com/questions/34178287/difference-between-objective-and-feval-in-xgboost
    
    
        
    # run cross validation
    myxgb_cv <- xgboost::xgb.cv(params=myxgb_params, my_iris_dmat, nrounds=1000, early_stopping_rounds = 10, nfold=5)
    
    
    
    # find best number of iterations from within the cross validation results
    best_numb_iter <- which( min(myxgb_cv$evaluation_log$test_merror_mean) == myxgb_cv$evaluation_log$test_merror_mean)[1]
    
        
    
    # run the model on all of "train" with same parameters as the crossvalidation
    myxgb_model <- xgboost::xgboost(
        params = myxgb_params,
        data = my_iris_dmat,
        nrounds = best_numb_iter,
        print_every_n = 1
    )

    
    # plot the importance
    myxgb_imp <- xgboost::xgb.importance(feature_names=colnames(my_iris_dmat), model=myxgb_model)
    xgb.ggplot.importance(myxgb_imp)
    xgb.ggplot.deepness(myxgb_model)
    ?xgb.train