One-Step Fuzzy Ensemble Clustering Method via Embedding Ground-Truth Cluster Number Graphs
Zekang Bian, Jia Qu, Zhaohong Deng, Shitong Wang · IEEE Transactions on Fuzzy Systems · 2025
In multisource clustering tasks, the number of clusters in each source or view may not align with the number of ground-truth clusters. Existing ensemble clustering methods face two notable challenges: (1) developing a new ensemble framework that yields a final clustering result matching the ground-truth cluster count and (2) revealing consistency among all base clustering results. To address these challenges, we propose a novel one-step fuzzy ensemble clustering method (OS-FECM) that incorporates ground-truth cluster number graphs. Initially, OS-FECM establishes a one-step fuzzy ensemble framework that directly integrates all base fuzzy clustering results (i.e., membership matrices) with varying cluster counts, thereby eliminating reliance on the CA matrix typical of existing two-step ensemble frameworks. Furthermore, we construct a ground-truth cluster number graph, which maps the number of clusters in each base clustering result to the ground-truth cluster count in the final ensemble result. This graph reveals the consistency among all base fuzzy clustering results and illustrates the relationships between clusters in the base results and the ground-truth clusters. It is then embedded into the corresponding base fuzzy clustering results to enhance the final ensemble result. Lastly, we employ an alternating optimization method alongside a weighting mechanism to derive the final ensemble clustering result and adaptively assign importance to each base clustering result. Experimental evaluations across various datasets demonstrate that OS-FECM achieves clustering performance that is at least comparable to, if not superior to, that of other comparative methods.