GBM: does using a bag fraction of one pose any problems?

Question

I am trying to fit a gbm model to some poisson distributed data and have run a cross validation scheme for some of gbm the parameters, including bag fraction. My results show that a bag fraction of one gives the best accuracy (in terms of poisson deviance). I have tried to read about bag fraction and understand it represents the fraction of the data used in growing the trees in subsequent iterations, meaning that my model is best when 100% of the data i used in growing the trees. I am however unsure if there are any concerns of using a bag fraction of one, and was hoping someone here could enlighten me.

Thank you in advance!

Answer 1

Having a bagging fraction of $1$: While usually a sub-optimal case that might hint to over-fitting, having a large bagging fraction (i.e. use most, or all, of available data being used in each iteration) is not catastrophic. Especially if our sample is not very large. What is important is that we have a robust and clearly defined way of testing the booster's performance. If we have a well-structured validation schema, that is relevant to our problem at hand, then a bagging fraction of $1$ is just one more hyper-parameter.

As a side-note not focusing on the function gbm exclusively: it has happened to me regularise excessively the overall fit through $L_1$ and $L_2$ regularisation terms. That lead to more iterations and higher bagging fractions being select. Some of the unusual hyper-parameter values might therefore be reflections of something more endemically wrong.

Answer 2

I have found that repeatability is better with a bag fraction of 0.8, compared to 0.95. Also, I found that altering the learning rate based on minimizing variability between runs was also achievable, at 0.00375 they ranged from 1000-1600, but at 0.0038 they ranged from 850-2100. I feel that it is more robust, if multiple runs are similar in their tree numbers, but I don't know any facts behind that.