The Online Customer: New Data Mining and Marketing Approaches By ...

Global model (no segmentation), k-means, the RFM-based (Recency, Frequency and Monetary value) segmentation model and GLIMMIX (Generalized Mixture Regression Model). With RFM, customers can be grouped according to how recently they bought from the retailer, how frequently they bought in a given time period, and how much they spent in that same period. RFM is a popular market segmentation model in the marketing literature and practice (Shepard 1995). Using the site-centric purchasing data we have, we are able to derive values for Recency, Frequency and Monerary value. Customers with the same value for RFM were grouped together. For example, customers with {Recency = low, Frequency = high, Monetary = high} were grouped into one segment. For every possible combination of the values of Recency, Frequency and Monerary value, there is a segment. In the prediction stage, a customer is assigned to one of the segments according to the values of the customer’s R, F and M. In our experiment, we have three possible values (low, medium, high) for R, F and M. See (Hartigan 1975) for the description of k-means and see Appendix 2 for the description of GLIMMIX. Table 4 presents the results comparing the predictive accuracy of the 5 approaches on the hold-out samples.

In our experiment, RFM was not significantly different from the Global model (p = 0.3478); GLIMMIX was significantly better than