Taking a machine learning system from prototype to production requires a lot more than what is typically covered in an introduction to machine learning. In this talk we show how NRK built a recommender system for on-demand video streaming, taking it from an internal prototype to a service running in production. TV viewing habits are changing, and NRK needs a recommender system to stay relevant and get better use out of NRK’s content catalog. In addition to explaining the algorithms we use, we will discuss challenges around data collection, infrastructure, model management and testing.
On to the buzzwords: The system is implemented in Scala using Apache Spark. Training runs in Google Cloud Dataproc with extensive use of Google BigQuery. Recommendations are served through a Play Framework API, deployed as Docker containers running on Mesos in both a public cloud (Microsoft Azure) and a private cloud.
Tune in to learn how you can make your system smarter by making better use of data you already have.