Whitrow et al. investigate the important problem of detecting fraudulent credit card transactions. A traditional approach to detecting a fraudulent transaction is based on a transaction-level classification. Another method is to consider behavioral patterns at the account level--some credit companies were relatively successful using this.

This paper proposes a methodology that is in line with transaction-based aggregation. The essence of the approach is to raise the alarm of a fraudulent transaction as soon as possible, by using accumulated information on a series of transactions for the account within a time window. The authors take this a step further by introducing a framework of transactional aggregations, so multiple classifiers can be adopted within, for example, support vector machines, k-nearest neighbors, random forests (which seem to perform better on the aggregated data), or logistic regression. There are limitations to the framework associated with aggregation window selection, but it does have some advantages over a transactional-level classification.

The power of aggregations within the time windows combined with a hidden Markov model could be especially promising; it appears to be a natural combination for the approach--aggregates drive state transitions, arriving ultimately at a healthy or fraudulent state (predominantly at the account, rather than the transaction, level). However, Whitrow et al. have not yet considered this option.