[Talks@IDSIA] Monday h 10:30 about probabilistic graphical models

Hi all, Hi all, next week two researchers from Seoul National University will visit us. We will have two talks to illustrate each other research activities. First talk will be at 10:30, second talk will start at about 11. They will be both in room 222. h10:30 == Supervised Belief Propagation: Supervised Inference on Complicated Graphs Mr. Jaemin Yoo, PhD student at the Seoul National University Abstract: Belief Propagation (BP) is a general graph inference algorithm for probabilistic graphical models. Because of its scalability and generality, BP has been widely used for solving real-world problems: malware detection, web classification, recommendation, etc. However, BP has a problem in modelling complicated structures because the most important parameter in the model, the propagation parameter, should be chosen manually. Although cross-validation nearly solves the problem for small models, BP couldn’t enlarge the domain of the algorithm because of this limitation. In this work, we propose Supervised Belief Propagation (SBP), a powerful inference algorithm which trains the propagation parameter using a gradient-based approach. As a result, it can model much complicated structures compared to BP including layered graphs, directed graphs, and featured graphs; none of these could be easily handled by BP. After that, we show SBP performs better than BP for all the existing problems, especially as models become complicated. Finally, we show applications of SBP for the sign prediction and node classification tasks. ==h 11:00 Learning Treewidth-Bounded Bayesian Networks with Thousands of Variables Mauro Scanagatta, PhD student at IDSIA. We present a method for learning treewidth-bounded Bayesian networks from data sets containing thousands of variables. Bounding the treewidth of a Bayesian network greatly reduces the complexity of inferences. Yet, being a global property of the graph, it considerably increases the difficulty of the learning process. Our novel algorithm scales both to large domains and to large treewidths, consistently outperforming the state of the art on experiments with up to ten thousands variables.