Résumé
Since most real-life data contain missing values, reasoning and learning with incomplete data has become crucial in data mining and machine learning. In particular, Bayesian networks are one machine learning technique that allows for reasoning with incomplete data, but training such networks on incomplete data may be a difficult task. Many methods were thus proposed to learn Bayesian network structure from incomplete data, based on multiple structure generation and scoring of their adequacy to the dataset. However, this kind of approaches may be time-consuming. Therefore we propose an efficient dependency analysis approach that uses a redefinition of probability calculation to take incomplete records into account while learning BN structure, without generating multiple possibilities. Some experiments on well-known benchmarks are described to show the validity of our proposal.