A Multi-objective Genetic Approach to Clustering with Solution Recommendation

Clustering, which attempts to partition data points into different clusters according to their features, is the most commonly used analysis method in data mining. Prototype-based clustering algorithms are widely used because of their simplicity and efficiency, but it is less effective for non-spherical distributions. Evolutionary algorithms are increasingly applied to improve the quality of partitioning solutions. However, firstly the existing prototype-based evolutionary clustering algorithms still have difficulty in giving unsupervised recommended solutions for complex distributions. Secondly, evolutionary clustering algorithms with complicated optimization process reduce the efficiency of the underlying clustering algorithm. Both of them seriously affect the general applicability of the prototype-based evolutionary clustering algorithms. In this paper, a novel prototype-based evolutionary clustering algorithm is proposed, which is valid for arbitrarily shaped datasets. In the proposed algorithm, a solution recommendation scheme that does not require a priori knowledge is developed and applied to preserve solution diversity and recommend a single solution. What’s more, a new cluster merging process is applied to promote population quality and accelerate convergence. To verify the effectiveness, experiments were conducted on different aspects of the proposed method using several artificial and real datasets. The results indicate that our algorithm is able to obtain suitable recommended solutions for spherical and non-spherical datasets and outperforms related clustering algorithms.