In machine learning there are tree principal areas where the computer is able to learn, the first one is supervised learning, the second one is unsupervised learning and the third one learning by reinforcement. On the first appears something that the literature calls the” learning problem” where they study the most common scenes in relationship to the data and the model, in this paper we purpose a modification of the MS-EM algorithm that estimated a hidden variable from incomplete data in a scene where the model is unknown, and this algorithm is able to get a Bayesian network that explains the presence and relationship of the hidden variable and the dataset.