Abstract
A novel approach based on supervised hierarchical clustering is developed with the purpose of discovering structure in data where labels are provided. Labels can come in the form of discrete-valued class labels or continuous-valued output variables to aid hierarchical clustering in discovering the structure and the number of clusters, in particular. In the proposed method, Clusters are linked together if their discrete-valued labels are the same or in the case of continuous output variables if their outputs are similar. Similarity within a cluster in the continuous case is expressed by a measure of internal cluster dispersion. Several experiments on synthetic data with discrete-valued class labels are conducted to demonstrate the algorithm's ability to discover class or data structure.