Counterfactual Evaluation of Slate Recommendations with Sequential Reward Interactions

Counterfactual Evaluation of Slate Recommendations with Sequential Reward Interactions (via Zoom)

Abstract: Users of music streaming, video streaming, news recommendation, and e-commerce services often engage with content in a sequential manner. Providing and evaluating good sequences of recommendations is therefore a central problem for these services. Prior reweighting-based counterfactual evaluation methods either suffer from high variance or make strong independence assumptions about rewards. We propose a new counterfactual estimator that allows for sequential interactions in the rewards with lower variance in an asymptotically unbiased manner. Our method uses graphical assumptions about the causal relationships of the slate to reweight the rewards in the logging policy in a way that approximates the expected sum of rewards under the target policy. Extensive experiments in simulation and on a live recommender system show that our approach outperforms existing methods in terms of bias and data efficiency for the sequential track recommendations problem.

Bio: James McInerney is a Senior Research Scientist at Netflix working on Bayesian approaches to machine learning, causality, and recommender systems. He was previously a researcher at Spotify and did postdoctoral work at Columbia University and Princeton University on scalable Bayesian inference. He has a PhD in Computer Science from the University of Southampton on the topic of machine learning for spatio-temporal data.