Bayesian Networks and Statistical Learning
Applications to complex system modelling
and diagnosis
Philippe LERAY, Olivier François, Ahmad Faour
contact: [email protected]
Structural learning – complete data
The DAG space has a super-exponential size → heuristics ! Constraint based methods (IC, PC, BN-PC. . .)
Score based methods
complete search in Tree space (MWST)
greedy search in DAG space, with node ordering (K2) or without (GS)
greedy search and Markov equivalence (GES)
Conferences : François & Leray RJCIA 03 (french), RFIA 04 (french) Journal : JEDAI 04 (french)
MWST = good performances vs. computation time MWST for GS initialisation = robust initialisation
Structural learning – incomplete data
Few methods deal with incomplete data
Usual principle = applying EM to score based methods greedy search in DAG space (SEM = GS+EM)
Conference : François & Leray EGC05 (french) [subm. to ECSQARU 05] :
MWST+EM = MWST + score estimation with EM
MWST+EM for SEM initialisation = robust initialisation
Perspectives :
greedy search and Markov equivalence = GES+EM constraint based methods and incomplete data
Structural learning – latent variables
Combinatorial explosion
Where are the latent variables in the DAG ? Cardinality ?
→ new operators in SEM
→ space restriction : hierarchical latent class model (HLC)
Conference : Leray & al. ECML03 Workshop (PGM for classification)
Tree augmented HLC
Perspectives :
SEM+EM = dealing with incomplete data and latent va-riable discovery
Structural learning – a priori knowledge
Using a priori knowledge to simplify the search space
Perspectives :
Dynamic bayesian networks (2TBN) = 2 structures : intra-slice (t) and inter-slice (t → t + 1)
Oriented object bayesian networks (OOBN), Multi-agent bayesian networks, ...
Complex system modelling and diagnosis
Discovering handwriting strategies of primary school children
I. Zaarour PhD thesis (completed in feb. 2004)
Collaboration with a psychology lab (PSY.CO Rouen)
Conferences : ECML03 Worshop - IGS03 - RFIA04 (french) Journal : IJPRAI 04
Complex system modelling and diagnosis
Intrusion detection in computer networks
A. Faour PhD thesis (begin sept. 2004)
Collaboration with a network security expert
Complex system modelling and diagnosis
Bayesian networks for classification
O. François PhD thesis (end envisaged in dec. 2005) Journal : RIA 2004 (french)
Dysfunction detection and localisation in a chemical reactor
Collaboration with a chemical process engineering lab (LRCP Rouen)
Conference : SFGP 2005 (french)
Micro-wave transistor thermical modelling
Project with Thales Air Defense and a aero-thermochemistry lab (CORIA Rouen) financed by Haute-Normandie Region
Scientific animation
French workshop on bayesian networks :
June 2001 – first workshop, Paris (co-organisation) March 2003 – second workshop, Rouen.
Jan. 2005 – French PGM workshop during EGC 2005 conference, Paris.
Software : BNT Toolbox for Matlab code contributions
responsable for structure learning package french BNT website and documentation
International activities
Members of PASCAL (Pattern Analysis, Statistical Modelling and Computational Learning) european network of excellence
Collaborations
Causal networks and structural learning – S. Meganck & B. Manderick, Computational Modeling Lab, Vrije
Selected bibliography (in english)
http ://asi.insa-rouen.fr/˜pleray/publisRB.php
International journals :
Zaarour, I. et al. (2004). Clustering and bayesian network approaches for
discovering handwriting strategies of primary school children. International Journal of Pattern Recognition and Artificial Intelligence, 18(7) :1233-1251.
International conferences :
Leray, P.et al. (2003). A bayesian model for discovering handwriting strategies of primary school children. In Working Notes of the Workshop on Probabilistic
Graphical Models for Classification, ECML/PKDD-2003, 49-57.
Zaarour, I.et al. (2003). A bayesian network model for discovering handwriting strategies of primary school children. In 11th Conference of the International Graphonomics society (IGS 2003), 178-181.
Misc :
Leray, P. and Francois, O. (2004). BNT structure learning package : Documentation and experiments. Technical report, Laboratoire PSI.
Selected bibliography (in french)
Books :
Naïm, P., Wuillemin, P.-H., Leray, P., Pourret, O., and Becker, A. (2004). Réseaux bayésiens. Eyrolles, Paris.
French journals :
Leray, P. and Francois, O. (2004). Réseaux bayésiens pour la classification -méthodologie et illustration dans le cadre du diagnostic médical. Revue
d’Intelligence Artificielle, 18/2004 :169-193.
François, O. and Leray, P. (2004). Etude comparative d’algorithmes
d’apprentissage de structure dans les réseaux bayésiens. Journal électronique d’intelligence artificielle, 5(39) :1-19.
French conferences :
Francois, O. and Leray, P. (2005). Apprentissage de structure dans les réseaux bayésiens et données incomplètes. In Proceedings of EGC 2005 (to appear), 1-6. Faour, A. and Leray, P. (2005). Réseaux bayésiens pour le filtrage d’alarmes dans les systèmes de détection d’intrusion. In Proceedings of EGC 2005 Atelier
Modèles graphiques probabilistes (to appear), 1-8.
Francois, O. and Leray, P. (2004). Evaluation d’algorithmes d’apprentissage de structure pour les réseaux bayésiens. In Proceedings of 14ème Congrès