Comparison of Unsupervised Classifiers.pdf

Comparison of Unsupervised ClassiersMuhammad Afzal Upal Eric M Neufeldupal@cs.ualberta.ca eric@cs.usask.caUniversity of Alberta University of SaskatchewanEdmonton, Canada. Saskatoon, Canada.Abstract: The activity of sorting like objects into classes without any help from anomniscient supervisor is known as unsupervised classication. In AI both symbolic andconnectionist camps study classication. The statistical classiers such as Autoclassand Snob search for the theory that can best explain the distribution of given data,whereas neural network classiers such as Kohonens networks and ART2 use the vectorquantization principle for classifying data. Previously, many studies have comparedsupervised classication algorithms, but the more challenging problem of comparingunsupervised classiers has largely been ignored. We performed an empirical comparisonof ART2, Autoclass and Snob. We highlight the strengths and weaknesses of the variousclassiers. Overall, statistical classiers, especially Snob, perform better than theirneural network counterpart ART2.Keywords: Unsupervised classication.Area of Interest: Concept formation and classication.1 IntroductionClassication is the process by which objectsare arranged into categories. Classicationcan be divided into two types: supervised andunsupervised. Supervised classication is aform of learning from examples. A supervisedclassier learns a function f when it is pro-vided with examples of the form (Xi; f(Xi))by a supervisor. Unsupervised classicationcan be thought of as learning from observa-tion. An unsupervised classier is providedwith some observations Xi, that have notbeen categorized by a supervisor and the goalis to look for similarities in the observed pat-terns and group them such that patterns inone group are similar to one another and dis-similar from the patterns in any other group.Unsupervised classication seeks a convenientand valid organization of data and not to es-tablish rules for separating future