Neighborhood information based semi-supervised fuzzy C-means employing feature-weight and cluster-weight learning

A semi-supervised fuzzy c-means algorithm uses auxiliary class distribution knowledge and fuzzy logic to handle semi-supervised clustering problems, named semi-supervised fuzzy c-means (SSFCM). Despite the performance enhancement obtained by adding additional information about data labels to the clustering process, semi-supervised fuzzy techniques still have several issues. All the data attributes in the feature space are assumed to have equal importance in the cluster formation, while some features may be more informative than others. Additionally, when the local information of samples is not considered through the cluster creation process, the SSFCM-based algorithms have low accuracy. In this research, a novel semi-supervised fuzzy c-means approach is proposed to handle the mentioned issues. The proposed approach is built on feature weighting, cluster weighting, and using neighborhood information. In this method, a novel fuzzy objective function based on feature weighting and cluster weighting is presented, which has two main parts: (1) a semi-supervised term that represents external class knowledge and (2) a spatial penalty term on the membership function that allows a clustering of sample to be affected by its neighbors. Both feature weights and cluster weights are determined adaptively through the clustering procedure. Combining both approaches forms an ideal clustering structure that is less sensitive to the initial centers. Furthermore, the penalty term acts as a regularizer and enhances the accuracy of the proposed approach. Experimental comparisons are conducted on several benchmark datasets to show the performance of the proposed approach. The experimental results further show that the proposed approach outperforms the state-of-the-art methods. The Matlab implementation source code of the proposed method is now publicly accessible at https://github.com/jafartanha/SSFCM-FWCW.