Decision Theory for Recommender Systems

Decision Theory for Recommender Systems (via Zoom)

Abstract: Decision theory allows a clear separation between decision making concerns, decision rules and utility functions with science i.e. states of nature and how likely they are to occur. We show that recommender systems can be modelled with decision theory where the decision rule is a simple ranker. Where utility is typically a successful individual recommendation and the model is complex involving difficult temporal dynamics. We demonstrate the power of decision theory for recommender systems by showing the semantic and methodological confusions that are present in many of the state-of-the art papers on the subject.

Bios: 

Flavian Vasile is the ML Recommendations Solutions Architect at Criteo AI Lab, with his main focus being on the development of Deep Learning-based Recommendation Systems and on introducing aspects of Causal Inference to Recommendation. His current research interests include Deep Sequential Models for Recommendation and understanding Recommendation as a decision-making system with reward uncertainty. Among his recent research publications, the work on Distributional Counterfactual Risk Minimization was accepted at AAAI 2019 and REVEAL 2020 and he is the co-organizer of the REVEAL Workshop series on Offline Evaluation and Bandit Learning for Recommender Systems in conjunction with the ACM RecSys.

David Rohde is the Lead of the Recommendation Research group at Criteo. His research interests are around Bayesian machine learning, offline evaluation and causal inference. He is one of the original creators of the RecoGym environment and has delivered several tutorials on counterfactual recommendation and publishes scientific papers in Bayesian statistics, recommender systems and causality.  He co-organises the Laplace's Demon webinar series on Bayesian machine learning at scale.