On the k-Medoids Model for Semi-supervised Clustering

Clustering is an automated and powerful technique for data analysis. It aims to divide a given set of data points into clusters which are homogeneous and/or well separated. A major challenge with clustering is to define an appropriate clustering criterion that can express a good separation of data into homogeneous groups such that the obtained clustering solution is meaningful and useful to the user. To circumvent this issue, it is suggested that the domain expert could provide background information about the dataset, which can be incorporated by a clustering algorithm in order to improve the solution. Performing clustering under this assumption is known as semi-supervised clustering. This work explores semi-supervised clustering through the k-medoids model. Results obtained by a Variable Neighborhood Search (VNS) heuristic show that the k-medoids model presents classification accuracy compared to the traditional k-means approach. Furthermore, the model demonstrates high flexibility and performance by combining kernel projections with pairwise constraints.