Semi-supervised K-Means Clustering by Optimizing Initial Cluster Centers

Semi-supervised clustering uses a small amount of labeled data to aid and bias the clustering of unlabeled data. This paper explores the usage of labeled data to generate and optimize initial cluster centers for k-means algorithm. It proposes a max-distance search approach in order to find some optimal initial cluster centers from unlabeled data, especially when labeled data can't provide enough initial cluster centers. Experimental results demonstrate the advantages of this method over standard random selection and partial random selection, in which some initial cluster centers come from labeled data while the other come from unlabeled data by random selection.