Mixture Models for Learning from Incomplete Data