Feature-weight and cluster-weight learning in fuzzy c-means method for semi-supervised clustering.

Semi-supervised clustering aims to guide the clustering by utilizing auxiliary information about the class labels. Among the semi-supervised clustering categories, the constraint-based approach uses the available pairwise constraints in some steps of the clustering procedure, usually by adding new terms to the objective function. Considering this category, Semi-supervised FCM (SSFCM) is a semi-supervised version of the fuzzy c -means algorithm, which takes advantage of fuzzy logic and auxiliary class distribution knowledge. Despite the performance enhancement caused by incorporating this extra knowledge in the clustering process, semi-supervised fuzzy approaches still suffer from some problems. 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. Thus the feature importance issue is not addressed in the semi-supervised category. This paper proposes a novel Semi-Supervised Fuzzy c -means approach, which is designed based on Feature-Weight, and Cluster-Weight learning, named SSFCM-FWCW. Inspired by the SSFCM, a fuzzy objective function is presented, which is composed of (1) a semi-supervised term representing the external class knowledge; (2) a feature weighting; and (3) a cluster weighting. Both feature weights and cluster weights are determined adaptively during the clustering. Considering these two techniques leads to insensitivity to the initial center selection, insensitivity to noise, and consequently helps to form an optimal clustering structure. Experimental comparisons are carried out on several benchmark datasets to evaluate the proposed approach's performance, and promising results are achieved. The Matlab implementation source code of the proposed method is publicly accessible at https://github.com/Amin-Golzari-Oskouei/SSFCM-FWCW. • This paper proposes a novel Semi-Supervised Fuzzy C-means approach based on Feature-Weight and Cluster-Weight learning. • An adaptive weight for each feature in each cluster applies based on its importance in forming the clusters. • The weight of the clusters is calculated dynamically to decrease cluster center initialization sensitivity. • Considering the conjunction of feature and cluster weighting helps form an optimal clustering structure. [ABSTRACT FROM AUTHOR]