Non-centroid-based discrete differential evolution for data clustering

Abstract

Data clustering can find similarities and hidden patterns within data. Given a predefined number of groups, most partitional clustering algorithms use representative centers to determine their corresponding clusters. These algorithms, such as K-means and optimization-based algorithms, create and update centroids to give (hyper) spherical shape clusters. This research proposes a non-centroid-based discrete differential evolution (NCDDE) algorithm to solve clustering problems and provide non-spherical shape clusters. The algorithm directs the population of discrete vectors to search for data group labels. It uses a novel discrete mutation strategy analogous to the continuous mutation in classical differential evolution. It also combines a sorting mutation to enhance convergence speed. The algorithm adaptively selects crossover rates in high and low ranges. We use the UCI datasets to compare the NCDDE with other continuous centroid-based algorithms by intra-cluster distance and clustering accuracy. The results show that NCDDE outperforms the compared algorithms overall by intra-cluster distance and achieves the best accuracy for several datasets.