This paper addresses the task of helping investigators identify characteristics in credit-card frauds, so as to establish fraud profiles. To do this, a clustering methodology based on the combination of an incremental variant of the linearised fuzzy c-medoids and a hierarchical clustering is proposed. This algorithm can process very large sets of heterogeneous data, i.e. described by both categorical and numeric features. The relevance of the proposed approach is illustrated on a real dataset containing next to one million fraudulent transactions.