Incremental machine learning at scale [Davorin Kopič, Zemanta]
Incremental learning is a way of updating machine learning models by constantly feeding them new data as it becomes available. It can be a great way of keeping ML models always up to date, even when you’re dealing with massive amounts of data.
It is a very powerful learning paradigm - but getting it right can be tricky. From choosing the right data flows and pipelines, to selecting appropriate ML algorithms and learning schemes, many pieces need to fall in place.
In this talk I will present how we productized incremental machine learning to make millions of predictions per second while being constantly up to date with an ever-changing environment; and share a couple of tips, tricks and shortcuts about how to productize incremental learning effectively.
At Zemanta, our algorithms participate in upwards of half a million auctions for online advertising space every second. Our machine learning models decide which ads to show to each user and for what price. All in real-time, before the webpage even loads. Online, where everything is changing and expanding constantly, you have no choice but to keep up - incrementally!
Davorin Kopič is the Head of Data Science at Zemanta, an Outbrain company. He and his team are responsible for developing machine learning algorithms for real-time bidding, advertising campaign optimisation automation, fraud detection, crunching and analysing TBs of data, ...
Zemanta is building the most advanced native advertising platform in the world. Marketing agencies use their platform to run native advertising campaigns that reach millions of people every day. In July 2017, Zemanta joined Outbrain, the world’s largest content discovery platform.