MUTE: Majority under-sampling technique

An application which operates on an imbalanced dataset loses its classification performance on a minority class, which is rare and important. There are a number of over-sampling techniques, which insert minority instances into a dataset, to adjust the class distribution. Unfortunately, these instanc...

Full description

Saved in:
Bibliographic Details
Main Authors: Bunkhumpornpat,C., Sinapiromsaran,K., Lursinsap,C.
Format: Conference or Workshop Item
Published: 2015
Online Access:http://www.scopus.com/inward/record.url?partnerID=HzOxMe3b&scp=84860630816&origin=inward
http://cmuir.cmu.ac.th/handle/6653943832/38626
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Chiang Mai University
Description
Summary:An application which operates on an imbalanced dataset loses its classification performance on a minority class, which is rare and important. There are a number of over-sampling techniques, which insert minority instances into a dataset, to adjust the class distribution. Unfortunately, these instances highly affect the computation of generating a classifier. In this paper, a new simple and effective under-sampling called MUTE is proposed. Its strategy is to get rid of noise majority instances which over-lap with minority instances. The removal majority instances are considered based on their safe levels relying on the Safe-Level-SMOTE concept. MUTE not only reduces the classifier construction time because of a downsizing dataset but also improves the prediction rate on a minority class. The experimental results show that MUTE improves F-measure by comparing to SMOTE techniques. © 2011 IEEE.