Abstract:

Semisupervised clustering employs a small amount of labeled data to aid unsupervised learning. For the poor ability of the clustering algorithm based on the K-means for nonspherical clusters problems, this paper presents a new idea that considers the influence of the labeled datapoints on the unlabeled datapoints in allocating category, puts forward a definition of gravitational influence degree between category and datapoint,and designs a semisupervised K-means clustering algorithm with a gravitational parameter.The experiments show that the new algorithm has better effect than the traditional semisupervised K-means clustering method in dealing with the distribution of nonspherical cluster clustering.

Key words: semisupervised clustering;constrainedKmeans;labeled data, gravitational influence;nonspherical cluster