This blog will contain a series of posts describing the use of the importance sampling estimator in the context of a recommender system. The first posts introduce the topic, and should be readable with only basic knowledge of probabilities theory. After that, I would like to explain some more advanced details on bias/variance tradeoff which typically arises when using this kind of estimator, and describe a variant we found useful in practice at Criteo.
Recommendation problem and Contextual bandits
This series of posts will describe the use of the importance sampling estimator in the context of a recommender system. In this first post, we will explain what is a recommender system and how to formalize it as an instance of a contextual bandit problem.
Counterfactual reasoning with the Importance Weighting Estimator
This post explains how to perform offline evaluation of a new version of the system using the “importance sampling estimator”