Fecha: Viernes, 27 de abril de 2012
Hora: 10:30 12:00 (nuevo horario)
Lugar: Salón de actos del instituto de investigación I3A, Albacete

ABSTRACT: 

The dynamic nature of data streams leads to a number of computational and mining challenges. In such environments, Bayesian network structure learning incrementally by revising existing structure could be an efficient way to save time and memory constraints. The local search methods for structure learning outperforms to deal with high dimensional domains. The major task in local search methods is to identify the local structure around the target variable i.e. parent children (PC). In this paper we transformed the local structure identification part of MMHC algorithm into an incremental fashion by using heuristics proposed by reducing the search space. We applied incremental hill-climbing to learn a set of candidate-parent-children (CPC) for a target variable. Experimental results and theoretical justification that demonstrate the feasibility of our approach are presented.

BIO:
Philippe Leray graduated from a French engineering school in Computer Sciences in 1993. He also got a Ph.D. (Computer Sciences) from the University Paris 6 in 1998, about the use of bayesian and neural
networks for complex system diagnosis. Since 2007, He is a full professor at Polytech'Nantes, a pluridisciplinary French Engineering University, teaching from basic statistics to probabilistic graphical models. He has been working more intensively on Bayesian Networks field for the past twelve years with interests for theory (bayesian network structure learning, causality) and application (reliability, intrusion detection, bio-informatics). Since January 2012, he is also the head  of the "Knowledge and Decision" research group, in the Nantes Computer Science lab. This research group is structured around three main research themes, data mining (association rule mining and clustering) and machine
learning (probabilistic graphical models), knowledge engineering, and knowledge visualization, with a transverse ambition, improving the performance, in terms of complexity but also "actionability" of mining and learning algorithms by integrating domain and/or user knowledge. https://sites.google.com/site/sitelinacod/


.