TR05-06 (2005) – A Comparative Study on Feature Selection Algorithms in Data Mining

Luminita Mihaela Moruz, Liviu Ciortuz

The field of feature selection in data mining hasexperienced a greatdevelopment in the past few years. Unfortunately,there are few worksthat show a comparative analysis of feature selectionalgorithmsintroduced until now.

In this technical report we analyze the effect thatfeature selectionhas on the performance of five learning algorithms. Weuse nine of themost known feature selection algorithms. We conductour study on twoclassification problems, Dermatology and Zoo, whosedatasets areprovided in the UCI (University of California atIrvine) repository.

We realize a comparative study on the effects that thechosen featureselection algorithms have on the performance and therunning time of thelearning algorithms used in solving the two problems.

On these problems, the feature selection algorithmsprovide an accuracywhich is 1.09% and respectively 1.98% higher than theaccuracy obtainedby learning algorithms alone. Additionally, thereduction of running timeespecially for time consuming learning algorithms likeSVMO is highlysignificant, 59% and respectively 91%.

Bibtex

@TechReport{csfsadm,
author = "Lumini{c t}a Mihaela Moruz and Liviu Ciortuz",
title = " A Comparative Study on Feature Selection Algorithms
           in Data Mining",
institution = "``Al.I.Cuza'' University of Ia{c s}i, 
                 Faculty of Computer Science",
year = "2005",
number = "TR 05-06",
note = "URL:http://www.infoiasi.ro/~tr/tr.pl.cgi"
}