Customer profile classification: To adapt classifiers or to relabel customer profiles?

Customer profiles are, by definition, made up of factual and transactional data. It is often the case that due to reasons such as high cost of data acquisition and/or protection, only the transactional data are available for data mining operations. Transactional data, however, tend to be highly sparse and skewed due to a large proportion of customers engaging in very few transactions. This can result in a bias in the prediction accuracy of classifiers built using them. The problem is even more so when identifying and classifying changing customer profiles whose classification may change either due to a concept drift or due to a change in buying behaviour. This paper presents a comparative investigation of 4 approaches for classifying dynamic customer profiles built using evolving transactional data over time. The changing class values of the customer profiles were analysed together with the challenging problem of deciding whether to change the class label or adapt the classifier. The results from the experiments we conducted on a highly sparse and skewed real-world transactional data show that adapting the classifiers leads to more stable classification of customer profiles in the shorter time windows; while relabelling the changed customer profile classes leads to more accurate and stable classification in the longer time windows.