Kernel-based multiobjective clustering algorithm with automatic attribute weighting

Clustering algorithms with attribute weighting have gained much attention during the last decade. However, they usually optimize a single-objective function that can be a limitation to cope with different kinds of data, especially those with non-hyper-spherical shapes and/or linearly non-separable patterns. In this paper, the multiobjective optimization approach is introduced into the kernel-based attribute-weighted clustering algorithm, in which two objective functions separately considering the intracluster compactness and intercluster separation are optimized simultaneously. Meanwhile, the sampling operation and efficient clustering ensemble method are incorporated with the projection similarity validity index approach to obtain the clustering solution, which can effectively reduce the computing time especially for large data. Experiments on many data sets demonstrate that, the proposed algorithm in general outperforms the existing attribute-weighted algorithms and the computing efficiency for selection of the final solution is improved by a large margin. Moreover, its merit in terms of the partition and cluster interpretation tools is shown.