英语论文网

different fromanyof rtþkj in Rtþk.We used the terms ‘similar’ and ‘quite different’ in theabove definitions. These terms are used to compare tworules in syntactic aspects and to judge their degree ofsimilarity and difference. But the terms ‘similar’ and ‘quitedifferent’ are subjective and different for each individual.Therefore we define the rule matching threshold (RMT)which can be differently determined by individual users.Finally, we define the degree of change as the measure ofhow much change has occurred. Evaluation of the degreeof change will be explained in the next section. Now, thechange detection problem is defined as follows using theabove definitions of each change type.Definition 4: Change detection problem The changedetection problemconsists of finding all emerging patterns,unexpected changes and added=perished rules betweendata sets which are collected from different periods andranking the changed rules in each type by the degree ofchange.
Expert Systems, September 2005, Vol. 22, No. 44. Methodology4.1. Overall procedureThe methodology for detecting the change in customerbehavior consists of the following three phases, asillustrated in Figure 2.In phase I, two rule sets are generated from each dataset by using decision tree analysis. For this purpose, wepresent two basic approaches to change mining in thedecision tree model: a new decision tree, and the sameattribute and same cut. Details are explained in thefollowing section.In phase II, the changed rule set is generated using therule matching method which compares two rules selectedfrom each rule set. We adapted the rule matching methoddeveloped by Liu and Hsu (1996) and modified it to detectall types of changed rules including emerging patterns,unexpected changes, and added and perished rules. Forefficient rule matching, similarity and difference measuresare developed.In phase III, various changed rules detected in phase IIare ranked according to the predefined degree of changethat is a measure of how much change has occurred. In thecase of emerging patterns, the growth rate or rate ofdecrease of the support value of the changed rule iscomputed and then the rules are sorted by the absolutevalue of the rate. The growth rate or rate of decrease playsthe role of the degree of change in an emerging pattern.Second, in the case of unexpected change, we adapted theunexpectedness concept from the study of Padmanabhanand Tuzhilin (1999). The degree of exception for certainexisting rules plays the role of degree of change in the caseof unexpected consequent change. Third, in the case ofadded or perished rules, we use the support value andmaximum similarity value of new or disappeared rules asthe degree of change. Maximum similarity value is definedin a later section, and more detailed explanations for themeasures and procedures are provided in the following.4.2. Discovery of rule change in the decision treeIn this paper, two approaches are used to mine changes inthe decision tree model: a new decision tree, and the sameattribute and same cut (Liu et al., 2000). The first approachmodifies the original tree structure, which makes thecomparison with the original structure difficult. In thesecond approach it is easy to compare the two decisiontrees, but the support values of the same rule of each treemight be different. Note that the basic decision treealgorithm we use in our study is C4.5. We have modifiedit in various places for change mining purposes.4.2.1. Same a本论文由英语论文网提供整理，提供论文代写，英语论文代写，代写论文，代写英语论文，代写留学生论文，代写英文论文，留学生论文代写相关核心关键词搜索。