This thesis addresses the problem of learning a probabilistic classification model in the presence of missing data and/orselection bias. This type of data is commonly encountered in many real-world applications (epidemiology, health, industry). To this aim, we propose to use directed acyclic graphs (DAGs) to models the mechanisms underlying the missing data and selection bias, based on the work of Judea Pearl. The purpose of the thesis is to establish certain conditions under which it is possible to learn a unbiased predictive model from these datasets. Specific case studies will be presented to validate our proposals.