Using More Initial Centers for the Seeding-Based Semi-Supervised K-Harmonic Means Clustering

Abstract

In the initialization of the traditional semi-supervised k-means, the mean of some labeled data belonging to one same class was regarded as one initial center and the number of the initial centers is equal to the number of clusters. However, this initialization method using a small amount of labeled data also called seeds which are not appropriate for the semi-supervised k-harmonic means clustering insensitive to the initial centers. In this paper, a novel semi-supervised k-harmonic means clustering is proposed. Some seeds with one same class are divided into several groups and the mean of all data is viewed as one initial center in every group. Therefore, the number of the initial centers is more than the number of clusters in the new method. To investigate the effectiveness of the approach, several experiments are done on three datasets. Experimental results show that the presented method can improve the clustering performance compared to other traditional semi-supervised clustering algorithms.