lightGBM bagging_fraction 的工作机制

lightGBM中，Bagging的工作机制是什么？

官方文档中对这2个参数的介绍：

bagging_fraction ?︎, default = 1.0, type = double, aliases: sub_row, subsample, bagging, constraints: 0.0<bagging_fraction<=1.0

• like feature_fraction, but this will randomly select part of data without resampling
• can be used to speed up training
• can be used to deal with over-fitting
• Note: to enable bagging, bagging_freq should be set to a non zero value as well

 bagging_freq ?︎, default = 0, type = int, aliases: subsample_freq

• frequency for bagging
• 0 means disable bagging; k means perform bagging at every k iteration
• Note: to enable bagging, bagging_fraction should be set to value smaller than 1.0 as well

在Stack Overflow中，有这样的答案：

Mykhailo Lisovyi :

The code executes what documentation says- it samples a subset of training examples of the size bagging_fraction * N_train_examples. And training of the i-th tree is performed on this subset. This sampling can be done for each tree (i.e. each iteration) or after each bagging_freq trees have been trained.

For example, bagging_fraction=0.5, bagging_freq=10 means that sampling of new 0.5*N_train_examples entries will happen every 10 iterations

大白话：

每迭代 bagging_freq轮，进行一次bagging_fraction