12:00 - 12:30 arrival and lunch served
12:30 - 13:30 Charles's talk
13:30 - 14:00 informal discussion
Many machine learning tasks require representing and learning affinity between pairs of items, where items may be of the same or different type. For example, in reinforcement learning the long-term reward Q(s,a) can be viewed as the degree of appropriateness of the action a in the state s. This talk will explore the implications of the bilinear representation Q(s,a) = s'Wa. I will explain how to learn the matrix W via linear regression with any desired loss function, regularized or not. Experiments with a dataset of pairs of members of the eHarmony dating service show that learning affinity in this way can outperform learning a Mahalanobis distance metric. (Learning Mahalanobis distance is essentially a special case of learning bilinear affinity.) I will also explain how to extend learning bilinear affinity to learning models with latent features for predicting links between nodes in social networks and other graphs. Experimental results show that a new algorithm based on this idea is the most accurate to date for link prediction.
Dr. Charles Elkan is a professor in the Department of Computer Science and Engineering at the University of California, San Diego. In 1998/99 he was Visiting Associate Professor at Harvard University. Dr. Elkan is known for his research in machine learning, data mining, and computational biology. For example, the MEME algorithm he developed with his Ph.D. student Tim Bailey has been used in over 2000 published research projects in biology. Dr. Elkan has won several best paper awards and data mining contests, and some of his graduate students have become leaders at companies including Google, Yahoo, Zillow, and IBM, while others have held faculty positions at Columbia University, the University of Washington, and other universities inside and outside the U.S.