Summit 2018: Beyond the point estimate: uncertainty in neural networks for recommendations / Yoel Zeldes

Recommender systems need to optimize a delicate balance between exploring new recommendations and exploiting informative ones. Up until this point recommender systems employed mostly bayesian prediction algorithms and utilized the inherent measures of uncertainty to optimize the crucial exploration/exploitation balance. While DNNs obtain state of the art results, they present specific challenges for recommender systems due to their lack of standardized uncertainty measures needed for most exploration strategies. In this talk I’ll cover how we overcome this challenge at Taboola and how we use uncertainty measures to our advantage when serving content recommendations billions of times a day. MP3