Join GitHub today
GitHub is home to over 28 million developers working together to host and review code, manage projects, and build software together.Sign up
[MRG+2] FIX enforce deterministic behaviour in BaseBagging #9723
What does this implement/fix? Explain your changes.
Enforce deterministic results when
Any other comments?
I wonder whether this in practice degrades the quality of the bagging.
Certainly, it needs a mention in what's new under changed models, and for the subtlety of the change to be described: "The base estimator may be presented with the same features in a different order to in previous versions; now the features will always be in input order."
My main worry is that using a mask rather than an array of indices only works if
indices does not have repeated values, does it not? This is not necessarily the case if bootstrap=True I think.
True, I did not see it at first. So we need to modify the
I am not really familiar with this part, but it seems that it should not change anything in the computation itself.
@lesteve am I correct and go forward or I missed something?
changed the title from
[MRG+1] FIX enforce deterministic behaviour in BaseBagging
[MRG] FIX enforce deterministic behaviour in BaseBagging
Oct 21, 2017
I think this solution is fine. However, I wanted to have you thoughts regarding a potential backward compatibility. I would consider a bug that
What are you thoughts on that, shall we raise anything regarding the change of behaviour?
It was also one of my question: #9723 (comment)
IMHO, I would call this a bug-fix since returning indices is what I would expect. Otherwise this is impossible to reproduce the training. In addition, this is still possible to get the samples by converting the indices to a mask.