Feature Selection For Human Recommenders

Abstract

Recommendation systems struggle to incorporate rich features, such as those derived from natural language and images. While humans can readily process this sort of information, they cannot not scale in the same way that statistical/ML models can. As a result, hybrid-algorithms that make recommendations based on the outputs of both computers and humans are becoming increasingly popular. This talk will explore novel methods for determining what features the human side of these systems should be processing. It will outline how experimental methods (borrowed from the behavioral sciences) can be used to this end, along with how the human recommendations may be improved as a result.