Learning from Rational(*) Behavior

Machine Learning is increasingly becoming a technology that directly interacts with human users. In fact, much of the Big Data we collect today are the decisions that people make when they use the systems we built. This is already evident in search engines, recommender systems, and electronic commerce, while other applications are likely to follow in the near future (e.g., autonomous robots, smart homes, games). In this talk, I argue that learning from human interactions requires learning algorithms that explicitly account for human behavior and motivation. Towards this goal, the talk explores how integrating microeconomic models of human behavior into the learning process leads to new learning models that no longer misrepresent the user as a “labeling subroutine”. This motivates an interesting area for theoretical, algorithmic, and applied machine learning research with connections to rational choice theory, econometrics, and behavioral economics.

(*) Restrictions apply. Some modeling required.

Thorsten Joachims