英语论文网

s intended task of detecting significantchanges. The data set was prepared from a Korean onlineshopping mall which sells various consumer goods. The dataset contains customer profiles and purchasing history suchas age, job, sex, address, registration year, cyber money,number of purchases, number of visits, payment methodduring one year, and total purchase amount. We preparedtwo data sets to detect significant changes of purchasingbehavior by the customers. The first data set containsprofiles and purchasing history information of certaincustomers who had bought more than one cosmetic from1 February 2000 to 30 June 2000. The second data setcontains the same information but includes customers whohad made an additional purchase of cosmetics from 1 July2000 to 5 January 2001. After preprocessing the data forcleansing and discretization, we built the decision tree. Thesplitting criterion is the w2 test. The procedure to detectchange is explained according to three categories.5.1. Same attribute and same cut at time tWe processed the data at time t for cleansing anddiscretization and built the decision tree and correspondingTable 1: Value of measure for each type of changeType of change Value of measure to classifyEmerging pattern dij ¼ 0; ðPk2Aijxijk > 0or yij > 0or ‘ij > 0ÞUnexpected consequent dij > 0; d0ijZRMTUnexpected condition dij < 0; d0ijZRMTAdded rule sj < RMTPerished rule si < RMTExpert Systems, September
rules with the SAS 8.0 enterprise miner program. And withthe same attribute and same cut, we built the decision treeand corresponding rules of the data set at time tþk. Then,we compared the rules of time t and tþk and discoveredthe emerging pattern, the unexpected changes and theadded=perished rules. We evaluated the changed rules tofind what rules had changed most. Figures 3 and 4 show thedecision trees of time t and time tþk. The emergingpatterns, unexpected changes and added=perished rules areshown in Table 2. Significant emerging patterns, unexpectedchanges and added=perished rules are summarizedin Tables 3, 4 and 5.Sales visitLow HighLowHighTotal59.3%40.7%100.0%12318452076LowHighTotal80.9%19.1%100.0%11252651390PaymentCash Card77.5%22.5%100.0%654190844Reserved MoneyHigh Low75.4%24.6%100.0%51016667685.7%14.3%100.0%1442416886.3%13.7%100.0%47175546LowHighTotal80.9%19.1%100.0%11252651390AgeTeen 40s70.0%30.0%100.0%7310LowHighTotal16.7%83.3%100.0%7537545020sLow HighLowHighTotalLowHighTotalLowHighTotalLowHighTotalLowHighTotalLowHighTotal29.2%70.8%100.0%71724LowHighTotal8.4%91.6%100.0%1718520210.6%89.4%100.0%24202226OrdCntLowHighTotalFigure 3: Decision tree of the data set at time t in the same attribute and the same cut.Sales visitLow HighLowHighTotal56.3%43.6%100.0%33492294594378.0%22.0%100.0%30388333871PaymentCash Card74.0%26.5%100.0%18316422473Reserved MoneyHigh Low63.1%36.9%100.0%698407110582.8%17.2%100.0%1133235136886.3%13.7%100.0%1207191139815.0%85.0%100.0%31117612072AgeTeen 40s46.5%53.5%100.0%27315813.5%86.5%100.0%2371303154020sLow HighLowHighTotal0%0%0%000LowHighTotal8.3%91.7%100.0%222248.3%91.7%100.0%22224OrdCntLowHighTotalLowHighTotalLowHighTotalLowHighTotalLowHighTotalLowHighTotalLowHighTotalLowHighTotalLowHighTotalFigure 4: Decision tree of the data set at time tþk in the same attribute and the same cut
Expert Systems, September 2005, Vol. 22, No. 4From changed rule (1) in Table 3, we can see the rapiddecrease of customers with a high purchasing rate