Eskil Forsell
2016-03-21 16:30:43 UTC
Hi all,
I'm an economist with a background in more regular inference but there's a
huge interest in the field on how we can draw useful lessons from Machine
Learning. I think Gilles re-posted a previous email to him but I hope it's
alright to take this opportunity to send a reminder and also attach a link
to the actual paper (see below).
One of the most interesting intersections between classical econometrics
and ML is outlined in a new paper on how to explicitly look for
heterogeneities in treatment effects of randomized interventions:
Athey, S., and G. Imbens. (2015) "Recursive Partitioning for Heterogeneous
Causal Effects" (http://arxiv.org/pdf/1504.01132.pdf).
The paper is still a working paper but the splitting criterion they outline
should be fairly straightforward to implement for someone with a good
understanding of how the tree module in scikit-learn works.
I'm really excited about the potential of this new criterion for policy
making and Randomized Controlled Trials. If anyone wants to hear more about
why I think it would be useful or discuss potential implementation in a
scikit-learn fork please send me an email.
All the best,
Eskil
I'm an economist with a background in more regular inference but there's a
huge interest in the field on how we can draw useful lessons from Machine
Learning. I think Gilles re-posted a previous email to him but I hope it's
alright to take this opportunity to send a reminder and also attach a link
to the actual paper (see below).
One of the most interesting intersections between classical econometrics
and ML is outlined in a new paper on how to explicitly look for
heterogeneities in treatment effects of randomized interventions:
Athey, S., and G. Imbens. (2015) "Recursive Partitioning for Heterogeneous
Causal Effects" (http://arxiv.org/pdf/1504.01132.pdf).
The paper is still a working paper but the splitting criterion they outline
should be fairly straightforward to implement for someone with a good
understanding of how the tree module in scikit-learn works.
I'm really excited about the potential of this new criterion for policy
making and Randomized Controlled Trials. If anyone wants to hear more about
why I think it would be useful or discuss potential implementation in a
scikit-learn fork please send me an email.
All the best,
Eskil