Off-policy evaluation is the art of guessing how well a new decision policy would do using only data collected under an old one, no live experiment allowed. In contextual bandits the workhorse estimators are importance sampling and its weighted cousin, each with an annoying flaw: plain importance sampling is unbiased but high-variance, while the weighted version is bounded but biased.

This paper proposes Kernel-WIS, which tries to keep the boundedness of weighted importance sampling and the linearity of the plain version. The authors show it is asymptotically consistent and report that it beats strong baselines empirically, especially when the behavior policy that generated the data is itself misspecified, which is the realistic case.

The abstract is short and leans on empirical claims, so check the paper for the estimator's definition, its assumptions, and where the advantage actually holds.