Including high-cardinality attributes in predictive models: A case study in churn prediction in the energy sector