ASSFCM: Adaptive Semi-Supervised Fuzzy C-Means
Abstract
Data stream scenarios characterized by partially labeled data and evolving distributions pose significant challenges for clustering and classification methods. In particular, existing semi-supervised fuzzy clustering approaches often struggle to appropriately adapt the number of clusters over time, especially in the presence of concept drift, class overlap, and the emergence or disappearance of classes. In this work, we propose an extension of the Dynamic Incremental Semi-Supervised Fuzzy C-Means algorithm, referred to as Adaptive Semi-Supervised Fuzzy C-Means (ASSFCM). The proposed method introduces an adaptive mechanism that combines cluster splitting, driven by reconstruction error, with a merging strategy based on centroid distance, enabling dynamic control of the clustering structure. The effectiveness of the approach is evaluated on synthetic data streams designed to simulate different types of concept drift and varying levels of label availability. Experimental results demonstrate that ASSFCM achieves highly competitive classification performance and an improved trade-off between reconstruction accuracy and clustering stability compared to its previous dynamic versions. These findings highlight the importance of adaptive cluster control in semi-supervised data stream scenarios.