http://lattes.cnpq.br/8769839698528730; SILVA, Luiz Antonio Pereira.
Résumé:
Bayesian networks can be constructed based on expert knowledge, historical data, or both. However, changes in the application domain, inaccuracies or high complexity in the collected information can result in Bayesian networks productions with low usability and/or precision. Faced with this problem, it is essential to improve the generated model as new knowledge is collected, continuously incorporating new knowledge to the existing one. In this work, two studies are carried out from two different perspectives in order to evaluate and better understand the use of incremental learning algorithms of Bayesian network structures in different contexts of use. In the first study, a systematic literature review is carried out in order to identify and evaluate solutions for the incremental learning of Bayesian network structures, as well as to delineate directions of new related research. In the second study, two of the solutions found are experimentally evaluated using real and synthetic data in order to test them in different contexts and compare their performances regarding the quality of the network learned. In the systematic review, most of the relevant literature studies are gathered and it is identified that the learning procedures of these solutions can be classified as refinement or adaptation, in which the main difference between them is in how they use the new knowledge acquired. It is possible to identify with the empirical evaluation that the incremental solutions analyzed produce results with scores identical to those generated by batch learning solutions, but differ in the generalization of new data. It is also noticed that the characteristics of the context and restriction factors applied by the algorithms interfere in the generalization quality of the networks. In general, it is concluded that the incremental learning algorithms of Bayesian networks can be considered an acceptable solution in restricted contexts of use.