Feature-weighted fuzzy clustering methods: An experimental review

Soft clustering, a widely utilized method in data analysis, offers a versatile and flexible strategy for grouping data points. Most soft clustering algorithms assume that all the features present in the feature space of a dataset are of equal importance and neglect their degree of informativeness or irrelevance. Distinguishing between the relative importance of features in providing an optimal clustering structure has become a very challenging task. Many feature weighting methods have been proposed to deal with this problem in the field of soft clustering, which can broadly categorized into six major types: feature reduction-based, entropy-based, variance-based, membership-based, optimization-based, and meta-heuristic-based. This paper comprehensively reviews the most significant fuzzy clustering algorithms that employ feature weighting techniques. A taxonomy of the feature weighting-based fuzzy clustering algorithms is presented. Furthermore, all state-of-the-art approaches are implemented in Python and compared in terms of clustering performance by conducting various experimental evaluation schemes. In this comprehensive experimental analysis, 26 state-of-the-art clustering algorithms are evaluated on two synthetic and 18 benchmark UCI datasets based on Accuracy (ACC), Normalized Mutual Information (NMI), Precision (PR), Recall (RE), F1, Silhouette (SI) and Davies-Bouldin (DB) evaluation criteria. Moreover, the significance of the experimental comparisons is examined using Friedman and Holm's post-hoc statistical tests. The experimental analysis demonstrates the superior performance of variance-based feature weighting algorithms in most datasets. All the tested algorithms are implemented in Python, and the related source codes are shared publicly at https://github.com/Amin-Golzari-Oskouei/FWSCA. © 2024 Elsevier B.V.