Learning simple and explainable preference models able to approximate human preferences is one of the key challenges of multicriteria decision support. In this paper we consider the Choquet integral as a general preference aggregation function and we study the problem of learning an instance, as simple as possible, of this general model to fit a database of preference examples. More precisely, we want to learn a sparse Mobius representation of the capacity defining the Choquet integral. To this end, we study an approach to sparse preference learning based on iterative re-weighted L2 regularization and dualization. We show how to implement this approach and we share the results of numerical tests performed on synthetic preference data. We discuss the benefit of this approach compared to L1 regularization.