Abstract

In off-policy statistical inference, the focus lies on learning from historical data generated by one policy while aiming to improve the estimation or decision-making process under a different policy. This is particularly relevant in scenarios where it's impractical or risky to collect new data under the target policy. Conformal inference is a statistical framework that focuses on providing reliable uncertainty estimates alongside predictions. The conformal confidence interval does not require the regression model $Y|X$
to be correct, which serves merely as a "working model." In contrast, traditional predictive interval estimation requires the conditional density function of $Y|X$
to be correct to achieve the specified coverage probability. For complex data structures, conformal inference offers an effective approach to predicting outcome $Y$ by leveraging cutting-edge machine learning techniques, although these techniques themselves may not always be perfect.

In this talk, we will study conformal predictive intervals by utilizing the mixture structure in the off-policy evaluation problem. We found a great efficiency gain if the "working mixture model" is close to the true one.