英语论文网

t al. (2001)developed a change detection procedure and a measurefor evaluating the amount of change based on associationrule mining. The measure for evaluating the amount ofchange due to Song et al. (2001) is adapted withmodification in this research. Change detection based onassociation rule mining can be useful to identify changes ofcustomer behavior in unstructured and ill-defined situationsbecause of the unsupervised learning feature ofassociation rule mining. However, decision tree analysis inchange detection problems can be used in more structuredArticle______________________________________
Expert Systems, September 2005, Vol. 22, No. 4
situations in which the manager has a specific researchquestion and it also detects the change in classificationcriteria in a dynamically changing environment.Changes detected by decision tree analysis are usefullyapplied to plan various niche-marketing campaigns. Forexample, in a shop, if a manager can find out that thecriteria of a certain customers’ group for choosing aproduct has changed from price to design, then he=she willmodify the existing merchandising strategy for such agroup of customers. The methodology suggested in thispaper detects changes automatically fromcustomer profilesand sales data at different periods of time. The mostcommon approach to discovering changes between twodata sets is to generate decision trees from each data set anddirectly compare the rules from the decision trees by rulematching. But this is not a simple process for the followingreasons. First, some rules cannot be easily compared due todifferent rule structures. Second, even with matched rules,it is difficult to know what kind of change and how muchchange has occurred. To simplify these difficulties, we firstdefine three types of changes as the emerging pattern, theunexpected change and the added=perished rule. Then wedevelop similarity and difference measures for rule matchingto detect all types of change from different timesnapshot data. Finally, the degree of change is evaluated todetect significantly changed rules.2. Background2.1. Decision treesClassification using decision trees can be used to extractmodels describing important data classes or to predictfuture data trends. The example of a classification modelusing decision trees is the bank loan model whichcharacterizes customers as either safe or risky. Classificationand prediction have numerous other applicationsincluding credit approval, medical diagnosis, performanceprediction and selective marketing. Data classification is atwo-step process as explained in Figure 1.In the first step, a model is built describing a predeterminedset of data classes or concepts. The model isconstructed by analyzing database tuples described byattributes. Each tuple is assumed to belong to a predefinedclass, as determined by one of the attributes called the classlabel attribute. Typically, the learned or trained model isrepresented in the formof classification rules, decision treesor mathematical formulae. For example, given a databaseof customer credit information, classification rules can belearned to identify customers as having either excellent orfair credit ratings (see Figure 1(a)). The rules can be used tocategorize future data samples, as well as provide a betterunderstanding of the database contents. In the second stepin Figure 1(b), the model is used to classify future datatuples or objects for which the class label is not known. Forexample, the classifi