Ensemble Learning: Better Predictions Through Diversity

In the ongoing Netflix Prize competition (netflixprize.com), participants are asked to write a program that can predict how a user will rate any movie. Many of the frontrunners’ programs are, in fact, collections of programs. Each program in a collection takes a different approach to this prediction problem. For a particular user and movie, the predictions of the individual programs are combined to make the final prediction. By using such ensembles, these participants have seen the accuracy of their programs increase.

This talk will provide an introduction to how recommendation systems have been built, and consider the ensemble approaches used in the Netflix Prize as evidence of the value of combining approaches.