Does Removal of Noisy Granulated Datasets Matter to Performance of Decision Tree Generation?
Tzu-Ming Kuo, H. F. Hung, Chien‐Hsing Wu · 2023
Existing studies have recognized the effect of noise of granulated datasets on classification performance. Whether this effect continues an aid in generating decisional rules for the tree-based learning models needs to be disclosed. This study conducts an experiment that investigates the effect of noisy data on rule generation performance (RGP). The unsupervised (equal-width interval, EWI) and 28 supervised (minimum description length, MDL) techniques were used to granulate datasets. The decision-tree based classification model that either included or did not include 24 EWI and 28 MDL noisy granulated datasets were used, followed by testing and comparison on classification accuracy and RGP. Main results are as follows. Removal of noisy granulated datasets with EWI is advantageous to decision tree generation when original classification accuracy (OCA) of datasets is higher than 90% or less than 70%, but not obvious between 70% and 90%. Contrariwise, those with MDL is neither highly related to improvement of generation rate nor simplicity for scales of both higher than 90% and less than 70%, but slightly related to those with OCA between 70% and 90%.