Couterfactual reasonning

Post date: Sep 28, 2016 4:02:02 PM

Notes on Bottou's paper. Discussion led by Krikamol.

Key equations are 4 and 5

We evaluate the counterfactual expected value by importance sampling

this is a direct consequence of the Markov factorization

Potential outcome of Rubins (treatment, variable, and covariate)

This is what people do in domain adaptation

In equation (5) L would be the loss function

The denominator needs to be non zero. Maybe not enough data to explore the support of the distribution.

Compute confidence interval

Upper bound the values