E-commerce websites, social networks and on-demand video platforms reach an ever growing number in the billions of users.
To maximize conversion rates or user engagement, they need to recommend the right items to the right users at the right time.
The problem is, given some information about the user and their previous behavior, how to choose efficiently something they will positively interact with among up to hundreds of millions of possibilities?
Such problems where only very incomplete and uncertain information is available are very well addressed by probabilistic approaches, which allow to rigorously quantify the uncertainty with probability distributions and to reason about the relevant variables in a mathematically grounded way.
The use of hardware accelerators allows to use more expressive models, which increase the quality of the recommendation, while keeping a low latency of a few tens of milliseconds.
This leads to increased sale volumes, while keeping the cost of ownership of the (cloud or on premise) infrastructure reasonable as more computations can be packed on fewer resources.