With "sequential" recommendations we refer to the problem where a system recommends various "things" to a person over time to achieve long-term objectives. In this talk I will present two domains that I have worked on and the research challenges and solutions. The first domain is a system at some web site that recommends various offers. For he offers to be successful the system should reason about successive tempting offers. The second domain is a points of interest recommendation (POI) system, where it recommends various locations for a person to visit in a city, or attractions in a theme park. These locations are recommended in a sequence, where the next POI follows naturally from the previous, while satisfying user preferences. For both domains we used Reinforcement learning algorithms. Some of the research challenges addressed were off-policy evaluation, building a recommendation system from "passive" data that do not contain past actions/recommendations. Inferring the user propensity to listen to a recommendation and optimizing for long-term recommendation while minimizing recommendation fatigue.
Georgios received his Ph.D. degree in computer science in 2002 from Michigan State University. From 2002 to 2004, he was a post-doctoral associate at the Computer Science and Artificial Intelligence Lab at M.I.T. In October 2004, he joined Intel Research as a research scientist and in July 2011, he joined Yahoo! Labs as a scientist. Finally, he joined Adobe Research as a senior analytics researcher in July 2012. Georgios' interests include scaling up computational models of learning and planning under uncertainty and their applications to the real world. His projects have evolved over time from building intelligent agents that interact with the world, such as robot navigation, to agents that interact one to one with people, such as a personal and physical math coin tutoring system, to large scale interactions, such as marketing and advertising agents that interact with millions of people. These interaction systems reason over the evolution of people's behaviors and guide them to achieve long-term goals.