NYC Machine Learning Meetup
Hosted by eBay – Dan Melamed give a machine learning talk titled:
“How To Learn From What Your Users Might Not See”
(Source: NYC ML Meetup)
This talk will focus on contextual bandits and their applications.
Prediction problems such as ad placement and product recommendation are typically solved by ML from records of user interactions. However, such records are inherently biased in a way that cannot be overcome by merely adding more data or more features. Naïve ML on such biased data can fail catastrophically.
In this tutorial, Dan will show how to learn from such data in a principled, efficient, and unbiased manner. The techniques that he will describe were largely responsible for a click-thru rate gain of over 25% on MSN.com. The same techniques can often be used as an exponentially cheaper and faster alternative to A/B tests.
Dan will also describe an open-source implementation of these techniques that has recently been released as part of the Microsoft Decision Service (aka.ms/mwt).
Dan Melamed has been working on big data since long before Google or Hadoop existed. He has held positions at various start-ups, at Microsoft Research, at AT&T Labs-Research, and on the CS faculty of New York University. Most of his research has focused on machine learning for natural language processing. His engineering work is illustrated by software in areas such as machine translation, gradient boosting, plagiarism detection, computational advertising, anomaly detection, and trainable virtual agents. He earned a $16 bounty from the inventor of C++ for finding errors in the main C++ reference book.