A semi-supervised agglomerative hierarchical clustering method based on dynamically updating constraints

Abstract

A semi-supervised agglomerative hierarchical clustering method based on dynamically updating constraints is proposing in this research. Following the existing semi-supervised clustering algorithm, this method uses the must-link and cannot-link constraints. Instead of using the idea that the instances with must-link constraints are pre-clustered before agglomerating with the others, this method employs a more general and reasonable process. Firstly, must-link and cannot-link constraints are expanded to compose a constraints closure. Then, a standard agglomeration instructed by cannot-link constraints is processed. During this procedure, the must-link and cannot-link are dynamically updated according to the intermediate clustering results. This updating process guarantees the validity of the final results. The fundamental advantage of this method is omitting the pre-clustering process of the instances with must-link constraints. This modification ensures that data points gain a more reasonable agglomeration order, which may result in a significant improvement on the clustering results. This research also introduces an implementation of this model based on Ward0s method, leading to the C-Ward algorithm. The experimental analyses on both Artificial simulated datasets and real world datasets show that this method is much better than the others.