A learning approach to hierarchical feature selection and aggregation for audio classification