The paper proposes a scalable incremental clustering algorithm to process heterogeneous data-streams, described by both categorical and numeric features, and its application to the domain of credit-card fraud analysis, to establish dynamic frauds profiles. The aim is to identify subgroups of frauds exhibiting similar properties and to study their temporal evolution and, in particular, the emergence of fraudster behaviours. The application to real data corresponding to a one year fraud stream highlights the relevance of the approach that leads to the identification of significant profiles.