[Colloquium] JOB TALK TODAY: Ben Marlin, UBC

[Katie Casey]

DEPARTMENT OF COMPUTER SCIENCE

UNIVERSITY OF CHICAGO

Date: Friday, January 21, 2011
Time: 2:30 p.m.
Place: Ryerson 251, 1100 E. 58th Street

----------------------------------------------

Speaker:		Ben Marlin

From:		University of British Columbia

Web page:	http://www.cs.ubc.ca/~bmarlin/

Title: 		Machine Learning, Collaborative Filtering and the Missing at
Random Assumption

Abstract:  	Model estimation, inference and prediction in the presence of incomplete data are pervasive problems in machine learning and statistical data analysis. In this talk, I will focus on collaborative filtering and recommender systems, which were recently popularized through the million dollar Netflix Prize. In recommender systems like Netflix, users provide ratings for items like movies and the goal is to produce personalized recommendations for each user. Typical online collections contain thousands to millions of items while typical users rate a small fraction of items resulting in massively incomplete data sets. Building on the theory of missing data due to Little and Rubin, I will present a line of research investigating the implicit mechanisms that people employ when selecting the items they rate in recommender systems. I will describe how the properties of these mechanisms can invalidate the common practice of ignoring missing data when estimating and evaluating collaborative filtering models. I will present a probabilistic framework for rating prediction that can mitigate these issues by explicitly modeling both the ratings and the underlying missing data mechanism, resulting in substantially increased prediction accuracy. This work received the best technical paper prize at the ACM Conference on Recommender Systems in 2009 and was recently selected for the best papers track at the 2011 International Joint Conference on Artificial Intelligence. I will conclude with a very brief overview of the other areas that I actively work in including Gaussian graphical models, unsupervised feature induction and non-likelihood-based model estimation.

Host: 		Pedro Felzenszwalb

Refreshments will be served following the talk at 3:30 in Ryerson 255.