Generation of incomplete test-data using bayesian networks

Olivier François (1) , Philippe Leray (2)

1 LITIS - Laboratoire d'Informatique, de Traitement de l'Information et des Systèmes
2 LINA - Laboratoire d'Informatique de Nantes Atlantique

We introduce a new method based on Bayesian Network formalism for automatically generating incomplete datasets. This method can either be configured randomly to generate various datasets with respect to a global percentage of missing data or manually in order to handle many parameters. [1] proposed three types of missing data : MCAR (missing completly at random), MAR (missing at random) and NMAR (not missing at random). The proposed approach can successfully generate all MCAR data mechanisms and most of MAR data mechanisms. NMAR data generation is very difficult to manage automatically but we propose some hints in order to cover some of the NMAR data situations.

Domains: Machine Learning [cs.LG]

Olivier François, Philippe Leray. Generation of incomplete test-data using bayesian networks. IJCNN, 2007, Orlando, United States. pp.2391-2396, ⟨10.1109/IJCNN.2007.4371332⟩. ⟨hal-00412939⟩