Improved semi-supervised point-prototype clustering algorithms
N. Tazi Labzour, Amine M. Bensaid, James C. Bezdek · 2002
In this paper we show that the semi-supervised point prototype clustering algorithm (ssPPC) has a defect. ssPPC consists of (1) using a point-prototype clustering algorithm (PPCA) to overpartition a test set X/sup u/; (2) assigning a physical class label to each of the resulting clusters based on training data; and (3) computing the class membership of each element x/sub k//sup u//spl isin/X/sup u/ in each physical class, based on x/sub k//sup u/'s memberships and the label assigned to each cluster. First, we show that ssPPC can produce degenerate partitions of X/sup u/. Then, we propose two alternative approaches that fix this defect and guarantee nondegenerate classes. We apply the improved algorithms to the Iris data, and show that their performance is superior to the ID3 decision tree and Quickpropagation neural network classifiers.