Gilles Louppe
2016-03-16 10:12:15 UTC
Hi Eskil,
(CC: the scikit-learn mailing list)
Unfortunately, I would not have time myself to implement this new
criterion. In any case, given the recent publication of this paper, I
dont think we would add it to the scikit-learn codebase. Our policy is
to only include time-tested algorithms. That being said, maybe
someone from the mailing list would be interested in helping you
implementing this criterion in a separate fork.
Best,
Gilles
On 14 March 2016 at 22:48, Eskil Forsell
(CC: the scikit-learn mailing list)
Unfortunately, I would not have time myself to implement this new
criterion. In any case, given the recent publication of this paper, I
dont think we would add it to the scikit-learn codebase. Our policy is
to only include time-tested algorithms. That being said, maybe
someone from the mailing list would be interested in helping you
implementing this criterion in a separate fork.
Best,
Gilles
On 14 March 2016 at 22:48, Eskil Forsell
Dear Gilles,
I'm writing to you as the first author of the tree module in scikit-learn to
gauge your interest in implementing a novel and really useful (at least for
policy oriented economists like me) splitting criterion.
I'm a PhD student in economics at Stockholm School of Economics and my
research and work focuses largely on evaluating policy by using randomised
controlled trials. There has recently been a lot of buzz in the field of
economics of the potential intersection of machine learning and causal
inference. Much of this buzz has been inspired by a paper outlining how to
use a splitting criteria tailored to the idea that the splits will later be
used as subpopulation for estimating treatment effects on a hold-out sample,
thus yielding correct standard errors. (I'm attaching the paper.)
The authors are working on implementing the criterion in R but haven't yet
released anything publicly. I really believe that this method of estimating
heterogenous causal effects will be extremely popular among empirical
economists and potentially be of great use to policy-makers who want to
figure out how interventions work differently depending on personal
characteristics.
I had a look at the criterion file but quickly realized that this wouldn't
be something I could implement myself. If you're interested I'd love to talk
more about it. If you're not interested, perhaps you could point me in the
direction of someone who might be and who'd have no problem of implementing
the criterion? Based on my understanding of the paper it's actually a quite
simple extension of the MSE criterion slightly complicated by the fact that
instead of raw means, we're using treatment effects (which crucially depend
on a treatment indicator variable).
All the best and hope to hear from you soon.
Regards,
Eskil
I'm writing to you as the first author of the tree module in scikit-learn to
gauge your interest in implementing a novel and really useful (at least for
policy oriented economists like me) splitting criterion.
I'm a PhD student in economics at Stockholm School of Economics and my
research and work focuses largely on evaluating policy by using randomised
controlled trials. There has recently been a lot of buzz in the field of
economics of the potential intersection of machine learning and causal
inference. Much of this buzz has been inspired by a paper outlining how to
use a splitting criteria tailored to the idea that the splits will later be
used as subpopulation for estimating treatment effects on a hold-out sample,
thus yielding correct standard errors. (I'm attaching the paper.)
The authors are working on implementing the criterion in R but haven't yet
released anything publicly. I really believe that this method of estimating
heterogenous causal effects will be extremely popular among empirical
economists and potentially be of great use to policy-makers who want to
figure out how interventions work differently depending on personal
characteristics.
I had a look at the criterion file but quickly realized that this wouldn't
be something I could implement myself. If you're interested I'd love to talk
more about it. If you're not interested, perhaps you could point me in the
direction of someone who might be and who'd have no problem of implementing
the criterion? Based on my understanding of the paper it's actually a quite
simple extension of the MSE criterion slightly complicated by the fact that
instead of raw means, we're using treatment effects (which crucially depend
on a treatment indicator variable).
All the best and hope to hear from you soon.
Regards,
Eskil