Title:	EXTRACTING RULES FROM TRAINED SELF-ORGANIZING MAPS
Author(s):	Christos Pateritsas , Stylianos Modes , Andreas Stafylopatis
ISBN:	978-972-8924-30-0
Editors:	Nuno Guimarães and Pedro Isaías
Year:	2007
Edition:	Single
Keywords:	Data mining, Self-Organizing Maps, rule extraction, classification.
Type:	Full Paper
First Page:	183
Last Page:	190
Language:	English
Cover:
Full Contents:	click to dowload
Paper Abstract:	Computational intelligence algorithms, like neural networks, constitute a main field of machine learning research of increasing usage as part of the data mining process. The goal of data mining is to extract knowledge from raw data expressed in useful and human understandable ways. While neural networks present various advantages suited for data mining tasks, in most cases they operate like black boxes and the extracted knowledge is hidden inside the trained structure of the networks. The Self-Organizing Map (SOM) is a neural network model following the unsupervised learning concept. The map consists of a grid of units that contain all the significant information of the data set, while eliminating noise data, outliers or data faults. The simplicity of the representation of the resulting mapping as well as its visualization capabilities are features that make Self-Organizing Maps suitable for data mining applications. Although Self-Organizing Maps provide visualization of the extracted knowledge, still this knowledge could be expressed in a symbolic way, such as rules, so that it can be even more easily understandable by human experts. Propositional IF-THEN type rules can be considered as one of the most comprehensible and usable forms of depicting knowledge. Combining these rules with the visualization capabilities of the SOM model can result to a very powerful data-mining tool. In this paper, we present methods for extracting rules from a trained SOM network. Four different methods are implemented, depending on the formation of boundaries or on the discovery of clusters, from which the rules will be generated. The rules aim at labeling the input data with the proper category. Even though the SOM network is trained in an unsupervised manner, the set of rules that can be extracted from the network presents satisfactory classification accuracy. This fact indicates that the extracted rules sufficiently describe the input data set.
	Go Back