In many decision-taking applications of machine learning, such as recommender systems, it is very natural to select actions based on the empirical estimates of their expected reward values: the highest scoring actions are prescribed. We outline two major defects of the current argmax-driven approaches, namely Covariate Shift and Optimizer's curse, make a quick outline of potential solutions and review some of our recent work on the subject, namely causal embeddings for recommendation and distributionally robust counterfactual risk minimization.
Joint work with Stephen Bonner. The paper is available here.
Back to the workshop page