Categorizing Web pages on the subject of neural networks

✍ Scribed by N Vlajic; H.C Card

• Publisher: Elsevier Science
• Year: 1998
• Tongue: English
• Weight: 178 KB
• Volume: 21
• Category: Article
• ISSN: 1084-8045
• DOI: 10.1006/jnca.1998.0066

✦ Synopsis

Most of the existing techniques for page classification on the World Wide Web are based on text only analysis. Recently, several hypertext clustering algorithms have been proposed. These provide promising results when the clustering is based on combined term-similarity and hyperlink-similarity measures. However, both the traditional and the advanced techniques require improvements in the term-or word-vector representation of Web pages, especially when applied to Web collections dealing with one or a few particular topics. In this work we introduce an autonomous agent for hypertext classification which is implemented in Java. This paper describes the development related to text-only analysis, including a modification of a well known rule for information retrieval, and the utilization of word correlation. The algorithm has been employed in clustering Web pages related to the subject of neural networks. The results are useful in arriving at an efficient term-vector representation, in order to achieve a rapid and appropriate clustering based on content of on-line documents. The term vectors derived using this algorithm have been classified using a modified adaptive resonance theory (ART) algorithm, an unsupervised learning method in artificial neural networks which is proven to provide very accurate and sophisticated clustering. Examples of the results are presented in the paper, suggesting several benefits of using the methods.