Learning with Humans in the Loop

Making sense of digital information is a growing problem in almost
every domain, ranging from scientists needing to stay current with new
research, to companies aiming to provide the best service for their
customers. What is common to such "big Data" problems is not only the
scale of the data, but also the complexity of the human processes that
continuously interact with digital environments and generate new data.

To address this problem, I will show how we can develop principled
learning approaches that explicitly model the process of continuously
learning with humans in the loop while improving system utility. As
one example, I will present the linear submodular bandits problem,
which jointly addresses the challenges of selecting optimally
diversified recommendations and balancing the exploration/exploitation
tradeoff when personalizing via user feedback. More generally, I will
show how to integrate the collection of training data with the user's
use of the system in a variety of applications, ranging from long-term
optimization of personalized recommender systems to disambiguation
within a single search session.
Yisong Yue is a postdoctoral researcher in the Machine Learning
Department and the iLab at Carnegie Mellon University. His research
interests lie primarily in machine learning approaches to structured
prediction and interactive systems, with an application focus in
problems pertaining information systems. He received a Ph.D. from
Cornell University and a B.S. from the University of Illinois at
Urbana-Champaign. He is the author of the SVM-map software package for
optimizing mean average precision using support vector machines. His
current research focuses on machine learning approaches to diversified
retrieval and interactive information retrieval.