Automatic determining optimal parameters in multi-kernel collaborative fuzzy clustering based on dimension constraint

Most cluster assignments using the traditional kernel clustering method strongly depend on the selection of the initial values. Under this scenario, directly using dimension reduction methods to deal with high-dimensional heterogeneous data in preprocessing might destroy the integrity of the original data. To overcome these limitations, this study constructs a high-dimensional kernel feature space that the high-dimensional heterogeneous data is mapped into this space. The proposed method implements kernel dimension constraint (KDC) on the mapped data, from which the data processed by the KDC forms the maximize hyper-plane data. Then we use the multi-kernel collaborative fuzzy clustering based on kernel learning to partition the maximize hyper-plane, called KL-MK-CoC, thus ensures the integrity of original data and also reduces the running time. It determines the optimal parameter in the multi-kernel clustering process, and we also embed the feature weight computation into the clustering procedure. Therefore, the combination of multiple kernels and the automatic adjustment of kernel weights renders the proposed algorithm more immune to unreliable features. Experimental results and comparisons demonstrate the excellent performance of KL-MK-CoC with its effectiveness in practice.