Active Semi-supervised Spectral Clustering Based on Pairwise Constraints
Xia Li · Dianzi xuebao · 2010
Semi-supervised clustering uses a small amount of supervised data such as pairwise constraints to aid unsupervised learning.The improved clustering performance depends heavily on the choice of constraints.This makes it important to explore the appropriate pairwise constraints for semi-supervised clustering.This paper presents a method for actively selecting informative pairwise constraints,which corresponds to pick up data pairs far apart in the same cluster and those close in different clusters.An active semi-supervised spectral clustering(ASSC) is then developed by utilizing the selected pairwise constraints to adjust the distance matrix in spectral clustering.As a result,the intra-cluster distance is decreased and the inter-cluster distance is increased.Experimental results on UCI benchmark data sets and artificial data set show that these informative pariwise constraints lead to substantial performance enhancement over the random selective pairwise constraints spectral clustering.