Integrating Feature and Instance Selection Techniques in Opinion Mining

Abstract

Opinion mining focuses on extracting polarity information from texts. For textual term representation,differentfeatureselectionmethods,e.g.termfrequency(TF)ortermfrequency– inverse document frequency (TF–IDF), can yield diverse numbers of text features. In text classification,however,aselectedtrainingsetmaycontainnoisydocuments(oroutliers),which candegrade theclassificationperformance.Tosolve thisproblem, instanceselectioncanbe adoptedtofilteroutunrepresentativetrainingdocuments.Therefore,thisarticleinvestigatesthe opinionminingperformanceassociatedwithfeatureandinstanceselectionstepssimultaneously. Two combination processes based on performing feature selection and instance selection in differentorders,werecompared.Specifically, twofeatureselectionmethods,namelyTFand TF–IDF, and two instance selection methods, namely DROP3 and IB3, were employed for comparison. The experimental results by using three Twitter datasets to develop sentiment classifiersshowedthatTF–IDFfollowedbyDROP3performsthebest. KeyWORDS Feature Selection, Instance Selection, Opinion Mining, Text Classification