Classification based on Predictive Association Rules

Abstract

Recentstudies indatamininghaveproposed a newclassification approach, called associative classification, which, according toseveral reports, suchas17,61,achieves higher classification accuracy thantraditional classification approaches suchasC4.5. However, theapproach alsosuffers fromtwomajordeficiencies: (1)itgenerates a verylarge numberofassociation rules, whichleads tohighprocessing overhead; and (2)itsconfidence-based ruleevaluation measure mayleadtooverfitting. Incomparison withassociative classification, traditional rule-based classifiers, suchasC4.5, FOILandRIPPER,are substantially faster buttheir accuracy, inmostcases, maynot beashigh. Inthispaper, we propose a newclassification approach, CLoPAR (Classification basedon Predictive Association Rules), whichcombines theadvantages ofboth associative classification and traditional rule-based classification. Instead ofgenerating a largenumberof candidate rulesas inassociative classification, CLoPAR adopts a greedy algorithm togenerate rulesdirectly from training data. Moreover, CLoPARgenerates andtests more rules thantraditional rule-based classifiers toavoidmissing important rules. Toavoid overfitting, CLoPARusesexpected accuracy toevaluate eachruleandusesthebestk rules in prediction. IndexTerms-association rule, rule-based classification