Using Random Effects to Account for High-Cardinality Categorical Features and Repeated Measures in Deep Neural Networks.