Shrinking Number of Clusters by Multi-Dimensional Scaling.
Tae-Chang Jee, Hyunjin Lee, Yillbyung Lee · Semantic Web Applications and Perspectives · 2006
Clustering is to divide given data and then, automatically find out the meanings hidden in the data. It analyzes data, which are difficult for people to check in detail, and then, makes several clusters consisting of data with similar characteristics. Clustering, which is used in various fields, is automatically done without human interference, but the number of clusters should be decided by men in advance. The number of clusters is a very important element because the result of clustering can be different, depending on the number of clusters. Therefore, this paper proposed a method of deciding the number of clusters, which is projecting the center of a cluster on the two-dimensional plane by use of Multi-Dimensional Scaling, and then, combining the clusters. As a result of experimenting this method with real data, it was found that clustering performance became better. in-depth analysis by laying up the documents with similar characteristics, not by accurately separating all the documents. Classification needs no more worrisome because the subjects to be classified are clear and the number of classifications is same as that of the subjects to be classified. Clustering allows a user to set up the number of clusters and its result has its own meaning. However, from the viewpoint of general users, it may be felt inconvenient that the result of clustering depends on the number of clusters and that the best optimal number of clusters can be obtained after users test with various numbers of clusters. Accordingly, a method to automatically decide the number of clusters is necessary, but there is still a lack of the researches compared with the researches of the clustering algorithm. People want to get a result in a very short time in information retrieval system. Besides the clustering performance is superior, it will be inconvenient system if the clustering takes too long time. Shrinking the number of clusters is aim to enhance the clustering performance, but if it takes too long time, it will be meaningless task after all. Therefore, this paper studied the method of automatically deciding the number of clusters without users' repetitive tests and without giving great influence on the clustering time. This paper is composed of: descriptions of the existing techniques of deciding the number of clusters (chapter 2); explanation of how to reduce the number of clusters (chapter 3); usefulness of the method proposed in chapter 3 through analysis of test results (chapter 4); and suggestions for future studies (chapter 5).