feature_weights参数没有影响Xgboost

Question

xgboost

有一个参数

feature_weights

应该会影响模型选择特征的概率，也就是说，我们可以给每个特征更多或更少的权重，但似乎该参数不起作用或者是我做错了什么吗？

X <- as.matrix(iris[,-5])
Y <- ifelse(iris$Species=="setosa", 1, 0)

library(xgboost)
dm1 <- xgb.DMatrix(X, label = Y)
#I set different probabilities for each feature
dm2 <- xgb.DMatrix(X, label = Y, feature_weights = c(1, 0, 0, 0.01))
params <- list(objective = "binary:logistic", eval_metric = "logloss")

set.seed(1)



xgb1 <- xgboost(data = dm1, params = params, nrounds = 10, print_every_n = 5)

[1] train-logloss:0.448305 
[6] train-logloss:0.090220 
[10]train-logloss:0.033148 



xgb2 <- xgboost(data = dm2, params = params, nrounds = 10, print_every_n = 5)

[1] train-logloss:0.448305 
[6] train-logloss:0.090220 
[10]train-logloss:0.033148 

但是模型的行为完全相同，似乎参数

feature_weights

被简单地忽略了

Answer 1

看来该参数仅在

colsample_by*

参数之一小于 1 时才起作用（默认情况下它们都是 1）。