Towards Recommender Engineering: Reproducible Research and Algorithm Personalities

March 5, 2014 - 4:30 pm to 5:30 pm

118 MacLean Hall, 2 West Washington St, Iowa City, IA

Michael Ekstrand, Ph.D. candidate in computer science at the University of Minnesota, is the featured speaker at the CS Colloquium this Wednesday, March 5, 2014. There will be a reception at 4:00 p.m. in Muhly Lounge prior to the colloquium event.

Recommender systems are a pervasive technology in the current generation of data-heavy applications and services. They apply machine learning, data mining, and information retrieval techniques to make (usually personalized) recommendations of movies to watch, items to purchase, people to talk to, and many other things. They are a valuable tool in helping users make sense of large-scale information spaces and find things of particular interest.

Unfortunately, while we have a great deal of evidence that they are effective at helping users, increasing sales, etc., there is much we do not know about how and why they work, and what makes one approach work better than another. My work focuses on enabling and conducting research on how recommendation techniques differ from each other. In the long term, I hope this work will enable new recommenders to be directly engineered for particular tasks based on well-understood principles rather than re-developed for each individual application.

In this talk, I will present my work towards this objective: building open-source tools to support recommender research, conducting offline experiments on algorithm behavior with public data sets, and an experiment on user perception of differences in recommender outputs with users of the MovieLens movie recommender.

Bio: Michael Ekstrand is a Ph.D candidate in computer science with GroupLens Research at the University of Minnesota.