Machine learning with high-cardinality categorical features in actuarial applications (Q6556598)

The study develops and extends a generalised linear mixed model in a deep learning framework and provides a novel generalised linear mixed model neural network (``GLMMNet'') approach to model high-cardinality categorical features. This model has a double benefit; first, by combining a deep neural network with the GLMM structure, it takes advantage of the predictive power of deep learning models as well as of the statistical strength of GLMMs. Moreover, the model is extremely flexible and suitable for actuarial applications. In fact, the GLMMNet is applied to a real insurance dataset and its performance is compared to other models considered throughout the paper.