Postponing the Evaluation of Attributes with a High Number of Boundary Points.pdfVIP

Department of Computer Science Series of Publications C Report C Postp oning the Evaluation of Attributes with a High Numb er of Boundary Points Tapio Elomaa and Juho Rousu University of Helsinki Finland Postp oning the Evaluation of Attributes with a High Numb er of Boundary Points Tapio Elomaa and Juho Rousu Department of Computer Science PO Box FIN University of Helsinki Finland Series of Publications C Rep ort C Helsinki May pages Abstract The eciency of the otherwise exp edient decision tree learning can b e im paired in pro cessing dataminingsized data if sup erlineartime pro cessing is required in attribute selection An example of such a technique is optimal multisplitting of numerical attributes Its eciency is hit hard even by a single troublesome attribute in the domain In this pap er we examine what are the eects of p ostp oning to a later stage of decision tree construction the evaluation of attributes that have a high numb er of b oundary p oints Analysis of numerical attribute partition ing shows that there is a direct connection b etween the ratio of the numb ers of b oundary p oints and training examples and the maximum go o dness score of the attribute Furthermore class distribution information from prepro cessing can b e applied to obtain tighter b ounds for a numerical attributes relevance in class prediction These analytical b ounds however are to o lo ose for practical purp oses We exp eriment with heuristic metho ds which p ostp one the evaluation of attributes that have a high numb er of attributes The results show