The International Computer Science Institute is pleased to present the following talk: Ensembles of classifiers presented by Alberto Suárez Universidad Autónoma de Madrid and ICSI Friday, April 06, 2007 1:30 p.m. ICSI Lecture Hall 1947 Center Street, Suite 600 Berkeley, CA Numerous experimental studies show that pooling the decisions of classifiers in an ensemble can lead to improvements in the generalization performance of weak learners. In the first part of this talk we present two novel algorithms to generate ensembles of classifiers. The first method takes advantage of the intrinsic variability of the iterative growing and pruning tree construction procedure devised by Gelfand et al. [Gelfand, S.B., Ravishankar, C.S. and Delp, E.J. (1991), "An Iterative Growing and Pruning Algorithm for Classification Tree Design", IEEE Transactions On Pattern Analysis and Machine Intelligence Vol. 13, no. 2 138-150]. The second procedure consists in generating classifiers of the same type built using perturbed versions of the training data. These perturbed data are obtained by injecting random uncorrelated noise in the class labels of the original training examples. The final classification is obtained by averaging over classifiers obtained with independent realizations of the noise. The second part of the talk is devoted to exploring the importance of the order of aggregation in bagging ensembles. We present several heuristic rules to order the classifiers in bagging ensembles prior to aggregation. Early stopping in the aggregation of the classifiers ordered with these rules allow the identification of subensembles that require less memory for storage, have a faster classification speed and can perform better than the original bagging ensemble. Furthermore, this ensemble pruning procedure does not seem to deteriorate the robust performance of bagging in noisy classification tasks. Speaker Bio: Alberto Suárez received the degree of Licenciado in Chemistry from the Universidad Autónoma de Madrid, Madrid, Spain, in 1988, and the Ph.D. degree in Physical Chemistry from the Massachusetts Institute of Technology (MIT), Cambridge, MA, in 1993. After holding postdoctoral positions at Stanford University (USA), at the Université Libre de Bruxelles (Belgium), as a research fellow financed by the European Commission within the program “Training and Mobility of Researchers”, and at the Katholieke Universiteit Leuven (Belgium), he is currently a professor in the Computer Science Department of the Universidad Autónoma de Madrid (Spain). He has worked on relaxation theory in condensed media, stochastic and thermodynamic theories of nonequilibrium systems, lattice-gas automata, and decision tree induction. His current research interests include machine learning, quantitative and computational finance, time series analysis and information processing in the presence of noise. To learn more about ICSI, please visit our website: http://www.icsi.berkeley.edu For directions to ICSI, see: http://www.icsi.berkeley.edu/about/location.html A list of other talks and special events at ICSI is available here: http://www.icsi.berkeley.edu/cgi-bin/events/index.pl ------------------------------ Leah Hitchcock ICSI Communications Coordinator leahh@icsi.berkeley.edu ICSI Seminars mailing list http://mailman.ICSI.Berkeley.EDU/mailman/listinfo/seminars-list