In the framework of missing data imputation, Rubin formalized three types of missing data mechanisms upon definition of a missing data indicator matrix pointing out the missing-ness in the data matrix, assigning it a random variable with a conditional probability distribution given the data matrix depending on unknown parameters. Within Rubin’s paradigm, missing data imputation can be understood as a model selection problem, such as to estimate the performance of different models in order to choose the best one which generate sample data. This paper formalizes a new missing data imputation paradigm. Within statistical learning paradigm, missing data imputation can be understood as a model assessment problem, whatever is the probability model underlying sample data the goal is to minimize its prediction error (generalization error) on new data.