15. May 2014, 16:15
Ernst-Abbe-Platz 2, seminar room 3423
Semi-supervised structured output learning using operator-valued kernels
Dr. Celine Brouard
(Department of Information and Computer Science, Aalto University, Finland)
Many real world applications involve objects with an explicit or implicit structure. For example, we may want to use as inputs or outputs structured data such as texts, images, or proteins and genes in computational biology. In the talk, I will speak about a novel framework for semi-supervised learning from both structured inputs and structured outputs. This approach addresses the structured output prediction problem as an output kernel learning problem. By using the kernel trick in the output space, the kernel approximation problem reduces to learning a function with values in a Hilbert space. We then use an operator-valued kernel-based regression approach to approximate this function. I will present the results obtained with this method on several link prediction problems and on a drug activity prediction problem.
Reference: Brouard, C., d’Alché-Buc, F. and Szafranski, M. Semi-supervised penalized output kernel regression for link prediction. In Proceedings of the 28th International Conference on Machine Learning (ICML), 2011.