Tree-based algorithms supports continuous target? about causalml HOT 3 CLOSED

The uplift forest algorithms don't currently support continuous target variables, and we have no immediate plans to implement such a feature. The main reason is that our testing hasn't showed a major difference between uplift forests and meta-learners in the binary case, and we don't have a particular reason to expect such a difference in the continuous case. With all that being said, we're happy to revisit this if some new results emerge.

from causalml.

@t-tte Thanks for the quick response!

what about the casual tree algorithm defined in these 2 papers.
https://arxiv.org/abs/1902.00087. “Learning triggers for heterogeneous treatment effects
https://arxiv.org/pdf/1504.01132.pdf Recursive partitioning for heterogeneous causal effects

Because I found the tree visualization very useful in practice, compared to the meta learning.

Basically for continuous target, it uses RMSE for treatment difference as cost function and penalize with leaf variance and generalization error between train and validation.

from causalml.

Thanks for the comment, @DSXiangLi. We have implementation of causal tree from the second paper by Athey and Imbens, but it's still in experimental and work in progress. We will keep you posted once it's ready for general use.

from causalml.