Active Clustering Ensemble with Subspace Learning and Constraint Propagation

Heterogeneous data analysis under scenarios with scarce supervised information is crucial for intelligent systems. Active clustering can enhance the quality of clustering by strategically querying specific pairwise constraints. However, the valuable constraint information presents significant challenges to current methodologies: (1) Unsupervised data processing cannot benefit from pairwise constraints. (2) The common query criteria result in a redundant set of constraints. (3) The model lacks the capability to engage with the temporary constraint set. In order to address the aforementioned issues, this paper proposes a novel active clustering ensemble with subspace learning and constraint propagation (ACE-SLCP). Specifically, constrained self-expressive BLS-autoencoder (CSEBLSAE) is designed to learn compressed feature subspaces based on the constraint matrix and data structure. Then, query criterion grounded on two key characteristics is proposed to produce elitism in constraint set. The dynamic selection strategy further optimizes the model performance. Extensive experiments conducted on a variety of real-world datasets have consistently demonstrated that ACE-SLCP outperforms the comparison methods of both semi-supervised and active ensemble clustering.