Tales of Bayesian Regression
In this talk, I will discuss few case studies for unusually accurate predictions with unstructured data. We develop and utilize the method of Bayesian Regression for what we call "Latent Source Model". The basic insight is quite simple. There are only few distinct ways in which phenomenon of interest occurs. Be it topics trending on Twitter, price variations in Bitcoin or choices consumers make in going to restaurants, shopping clothes or watching movies. The method is embarrassingly parallel and scalable while being highly accurate.
Devavrat Shah is an Associate Professor with the department of Electrical Engineering and Computer Science at MIT. He is a Co-Founder and Chief Scientist of Celect, which helps retailers decide what to put where by accurately predicting customer choice using omni-channel data. His primary research interest is in developing large-scale machine learning algorithms for massive unstructured data. He has made contributions to development of “gossip” protocols and “message-passing” algorithms which have been pillar of modern distributed data processing systems.
He has received 2010 Erlang Prize from INFORMS. He is a distinguished young alumni of his alma mater IIT Bombay.