patterns generated from other clusters. Customers’ behavioral patterns can have different representations. A behavioral pattern can be represented as an IF-THEN rule. It can be represented by a collection of things (called an “itemset” in the data mining literature) a customer does together. It is also common to use sequences for pattern representation. Correlation between the values of different variables can also be considered as a type of pattern. We suggest that different domains may have different representations for what “patterns” are and for how to define differences between sets of patterns. We investigate the utility of pattern-based clustering for grouping Web transactions. In particular, we argue that itemsets are a natural representation for patterns in Web transactions and present GHIC (Greedy Hierarchical Itemset-based Clustering) – a pattern-based clustering algorithm for domains in which itemsets are the natural pattern representation.
After evaluating GHIC on 80 sub-datasets generated from a Web browsing data set, we further develop a modeling framework for building segment-level predictive models based on the pattern-based clustering approach and signature discovery techniques. Each category/cluster of customer transactions discovered by the pattern-based clustering approach can be characterized by its own distinguishing patterns. After we elicit multiple categories of customer transactions, we build one signature capturing the salient behavioral patterns for each category, as well as one predictive model for each category. In the prediction stage, a new transaction is compared with all the signatures and the closest signature is chosen. Then this new transaction is assigned to the category of transactions that the signature represents and the model associated with this signature is used to predict this transaction (or the models combined using a weighting scheme). Experiments conducted using online purchasing data are used in this study to evaluate the modeling technique and compare the proposed approach with other approaches from data mining and marketing (RFM, GLIMMIX, and k-means).