| Chapter 2: | Segmenting Customer Transactions Using a Pattern-Based Clustering Approach |
This is a limited free preview of this book. Please buy full access.
(Agrawal et al. 1995)) which reflects what a customer does together to represent certain behavioral patterns. For example, such a pattern could be that customer Y visits yahoo.com, google.com and excite.com in the same Web browsing session. It is also common to use sequences for pattern representation. One such example is that customer W visits cnn.com after msn.com. Correlation between the values of different variables can also be considered as a type of pattern. For example, the price of stock A is positively correlated with the price of stock B.
We suggest that different domains may have different representations for what “patterns” are and for how to define differences between sets of patterns. In the above consultant example, rules are an effective representation for patterns generated from the wireless data; however, in a different domain, such as time series data on stock prices, representations for patterns may be based on “shapes” in the time series. It is easy to see that traditional distance-based clustering techniques and mixture models are not well suited to discover clusters for which the fundamental characterization is a set of patterns such as the ones above.
One reason that pattern-based clustering techniques can generate natural clusters from customer transactions is that such transactions often have natural categories that are not directly observable from the data. For example, Web transactions may be for work, for entertainment, shopping for self, shopping for gifts, transactions made while in a happy mood and so forth. Although customers obviously do not indicate which situation they are in before starting a transaction, the set of patterns corresponding to transactions in each category will be different. Transactions at work may be quicker and more focused while transactions for entertainment may be long and across a broader set of sites. Hence, grouping transactions such that the patterns generated from each cluster are “very different” from those generated from another cluster may be an effective method to learn the natural categorizations.
This argument suggests the natural evaluation technique that is used in this essay. We combine transactions with a known natural category – Web