Circle-Clustering: A New Heuristic Partitioning Method for the Clustering Problem

This paper present a novel method to perform clustering of time-series and static data. The method, named Circle-Clustering (CirCle), could be classified as a partition method that uses criteria from SVM and hierarchical methods to perform a better clustering. Different heuristic clustering techniques were tested against the CirCle method by using data sets from UCI Machine Learning Repository. In all tests, CirCle obtained good results and outperformed most of clustering techniques considered in this work. In addition, CirCle was tested against others heuristic techniques considering time-series data from electric feeders in Santiago, Chile's capital city. The optimal solution of the min-cut clustering optimization problem was solved in order to identify the optimal solution for 883 datasets. The results show that the proposed method obtains an average of 81% of well-classified samples in all datasets. Also, as compared to other algorithms, CirCle made a better classification in 98.7% of the datasets as compared to the Model-Base Best BIC. As compared to K-means, Robust K-means and Ward's methods the new algorithm classified better in nearly 68% of the datasets.