A Comparison of Noise Handling Techniques

Imperfectionsin datacanarise frommanysources.Thequality of the datais of primeconcernto anytask that involves dataanalysis.It is crucialthat wehavea goodunderstanding of dataimperfectionsandthe effects of variousnoisehandlingtechniques.Westudyhere a numberof noise handling approaches,namely,robustalgorithmsthat are tolerant of someamountof noisein the data, filtering that eliminatesthe noisyinstancesfromthe input, andpolishingwhichcorrects the noisyinstancesrather than removingthem.Weevaluated the performanceof these approachesexperimentally.Theresults indicatedthat in additionto the traditionalapproachof avoidingoverfitting, bothfiltering andpolishingcanbeviablemechanismsfor reducingthe negativeeffects of noise. Polishingin particular showedsignificant improvementover the othertwoapproachesin manycases,suggestingthat even thoughnoisecorrectionaddsconsiderablecomplexityto the task, it also recoversinformationnotavailablewiththe other twoapproaches.

• Type: Conference
• Year: 2001
• Where: FLAIRS
• Authors: Choh-Man Teng
• Keywords: Artificial Intelligence, FLAIRS 2001, Namely, Noise Handling Approaches, Robustalgorithmsthat