Performance-enhanced rough $$k$$-means clustering algorithm

Abstract

Customer segmentation (CS) is the most critical application in the field of customer relationship management that primarily depends on clustering algorithms. Rough k-means (RKM) clustering algorithm is widely adopted in the literature for achieving CS objective. However, the RKM has certain limitations that prevent its successful application to CS. First, it is sensitive to random initial cluster centers. Second, it uses default values for parameters $$w_{l}$$ and $$w_{u}$$ used in calculating cluster centers. To address these limitations, a new initialization method is proposed in this study. The proposed initialization mitigates the problems associated with the random choice of initial cluster centers to achieve stable clustering results. A weight optimization scheme for $$w_{l}$$ and $$w_{u}$$ is proposed in this study. This scheme helps to estimate suitable weights for $$w_{l}$$ and $$w_{u}$$ by counting the number of data points present in clusters. Extensive experiments were carried out by using several benchmark datasets to assess the performance of these proposed methods in comparison with the existing algorithm. The results reveal that the proposed methods have improved the performance of the RKM algorithm, which is validated by the evaluation metrics, namely convergence speed, clustering accuracy, Davies–Bouldin (DB) index, within/total (W/T) clustering error index and statistical significance $$t$$ test. Further, the results are compared with other promising clustering algorithms to show its advantage. A CS framework that shows the utility of these proposed methods in the application domain is also proposed. Finally, it is demonstrated through a case study in a retail supermarket.