Semi-supervised clustering, methods and applications: A systematic review

Semi-Supervised Clustering (SSC) combines unlabeled data with limited prior information, such as labels, pairwise constraints, or domain rules, to guide the discovery of meaningful data partitions. Although SSC has been widely studied, recent advances have expanded the field beyond traditional must-link and cannot-link constraints, incorporating active learning, weak and distant supervision, label propagation, deep representation learning, and domain-specific constraint generation. This paper aims to present a systematic review on the several applications and worthwhile works that have been developed in SSC between 2004 and 2026, but also to provide an overview of the new processes of building the constraints. The reviewed methods are organized into a taxonomy according to their implementation strategy, type of prior information, constraint-generation process, scalability, robustness, and application domain. The analysis shows a clear evolution from classical constrained partitional methods toward hybrid frameworks that integrate representation learning, active querying, pseudo-labeling, and automated supervision. The results also reveal persistent challenges, including noisy or unreliable constraints, annotation cost, computational complexity, lack of standardized benchmarks, limited reproducibility, and difficulty comparing methods evaluated under different supervision budgets. By synthesizing algorithmic trends, supervision strategies, limitations, and open research directions, this review provides a structured overview of the current state of SSC and highlights opportunities for more scalable, robust, and interpretable clustering systems. • Semi-Supervised Clustering has significantly evolved from basic modifications of classical algorithms into complex, hybrid systems and deep learning frameworks that learn feature representations and cluster assignments end-to-end. • Methods for acquiring supervision have matured beyond passive constraint sets to include dynamic and automated strategies. • There is a lack of standardized benchmarks and inconsistent complexity reporting, highlighting a critical need for reproducible evaluation protocols and comprehensive comparative experiments.