Les modèles probabilistes relationnels : apprentissage et évaluation
Mouna Ben Ishak · HAL (Le Centre pour la Communication Scientifique Directe) · 2015
Statistical relational learning (SRL) appeared in the early 2000s as a new field of machine learning that enables effective and robust reasoning about relational data structures. Several conventional data mining methods have been adapted for direct application to relational data representation. Relational Bayesian Networks (RBNs) extend Bayesian networks (BNs) to a relational data mining context. To use this model, it is first necessary to build it: the structure and parameters of a RBN must be set manually or learned from a relational observational dataset. Learning the structure remains the most complicated issue as it is a NP-hard problem. Existing approaches for RBNs structure learning are inspired from classical methods of learning the structure of BNs. The evaluation of learning approaches requires testing datasets and evaluation measurements. For BNs, datasets are usually sampled from real known networks. Otherwise, processes to randomly generate the model and the data are already established. Both practices are almost absent for RBR. Moreover, metrics to evaluate a RBN structure learning algorithm are not yet proposed.This thesis provides two major contributions. I) A synthetic approach allowing to generate random RBNs from scratch. The proposed method allows to generate RBNs as well as synthetic relational data from a randomly generated relational schema and a random set of probabilistic dependencies. Also, we discuss the adaptation of the evaluation metrics of BNs structure learning algorithms to the relational context and we propose new relational evaluation measurements. II) A hybrid approach for RBNs structure learning. This approach presents an extension of the MMHC algorithm in the relational context. We present an experimental study to compare this new learning algorithm with the state-of-the-art approaches.