A couple of weeks ago, Criteo attended the ACM Recsys 2015 conference at the Technische Univerität Wien, Vienna.
Criteo makes display perform better : display the right ad, at the right moment, to the right user. So Romain Lerallut and Diane Gasselin gave a talk about the large scale real-time product recommender we use: what we do, how we do it, and what are the new challenges.
We must process recommendation requests in less than 100ms while we juggle with 1B users, 3B products large data sources and 20B/day of display data. In order to scale and be reactive, all the products cannot be filtered and ranked online. To address this issue, we pre-select sources of product offline using collaborative-filtering and logistic regression scoring and we then refine the score with the ad context online.
We are also facing various new challenges: offline ab-testing, longer-term user profile, NLP features, instant updates, full banner scoring, …
by Diane Gasselin for the Criteo R&D team.