陈国青 电子商务商务智能前沿课题(厦大2010)培训课件.ppt

* * ————— 2010.11 清华大学经济管理学院 陈国青 谢谢！ * 附：关联分类解析 －GARC/GEAR 传统方法：CBA 关联规则X?Y 特例：分类规则X?C （如：“居住A区/白领/频繁客户=银卡”） Apriori算法 其它主流方法：Decision-tree, NN, SVM,… 主要测度：可理解性、精度、效率 CBA：＋：可理解性（规则形式），精度 －：规则数目（大、冗余、冲突） GARC 《Decision Support Systems》 （Chen et al., 2006） * GARC 基本思想 引入信息熵，用于对1-itemsets确定最佳分裂属性，并被后面多项集包含 保存不满足支持度的集合－排除集，用于后面候选集的生成 确定若干规则冗余及冲突剪枝性质，并将它们融于频繁集生成过程 － 候选/频繁集个数大大减少 － 消除若干冗余和冲突 － 分类精度良好（与其它方法比较） * GARC细节 Compact ruleset/性质 check each rule XY?Ci in ?c, if there exists a rule X?Ci in ?c, then delete XY?Ci. That is, ?c = ?c ? {XY?Ci} check each rule X?Ci in ?c, if there exists a rule X?Cj in ?c that does not precede X?Ci, then delete X?Cj. That is, ?c = ?c ? {X?Cj}. check each rule XY?Cj in ?c, if there exists a rule X?Ci in ?c that precedes it, then delete XY?Cj. That is, ?c = ?c ? {XY?Cj}. Theorem 2.1: if b 50%, then rules X?Ci and X?Cj will not hold in T simultaneously. Theorem 2.2: Suppose rule X?Ci holds in T, (1) if 1 – Dconf(X?Ci) ?, then any itemset like XYCj is an excluded itemset; (2) if ||X|| 2|T|?, then any itemset like XYCj is an excluded itemset; where Y ? X = ?, Y ? Ck = ?, k = 1, 2, …,g, and g ? 2. * GARC细节 算法 1.???????? rule = { r | r is an 1-itemset, Dsupp(r) ? ? and Dconf(r) ? ?}; //initiating the set of qualified itemsets // 2.???????? excluded ={ e | e is an 1-itemset, and Dsupp(e) ?}; //initiating the set of excluded itemsets// 3.???????? bestattr = gain; 4.???????? if (??0.5) and (?r:X?Ci ? rule, ?r’:X?Cj ? rule such that r’ does not precede r) then 5.???????? rule = rule ? {X?Cj}; //deleting conflicting rules// 6.???????? for k from 2 to m do // m is the number of ancetedent attributes// 7.???????? empty(cand); // emptying cand, the set of candidate itemsets// 8.???????? if coverall(rule) 9.???????? break; 10.?????? for each transaction t in T do 11.?????? Ct = CandidateGen(t, bestattr, k); 12.??????