DCR: Discretization using Class Information to ReduceNumber of Intervals

Quality issues, measures of interestingness and evaluation of data mining models Workshop

图书情报档案学;计算机科学

Prachya Pongaksorn ; Thanawin Rakthanmanon ; Kitsana Waiyamai

PID  :  84236

来源: CEUR

【 摘 要 】

Discretization techniques for data set features have received increasing research attention. Results using discretized features are usuallymore compact, shorter, and accurate than using continuous values. In this paper,an algorithm called Discretization using Class information to Reduce number ofintervals (DCR) is proposed. DCR uses both class information and orderbetween attributes to determine the discretization scheme with minimumnumber of intervals. Attribute discretization order is determined based oninformation gain of each attribute with respect to the class attribute. Thenumber of intervals is reduced by deleting training data at each step of attributediscretization. Experiments are performed to compare the predictive accuracyand execution time of this algorithm with several wellknown algorithms.Results show that discretized features generated by the DCR algorithm containa smaller number of intervals than other supervised algorithms using lessexecution time, and the predictive accuracy is as high or higher.