A Unified Ensemble Clustering Framework via Higher-Order Graph Diffusion and Feedback-Based Refinement

Faria Chowdhury Mumu, Kife I. Bin Iqbal, Liang Du

Abstract: Ensemble clustering has emerged as a powerful paradigm to improve clustering robustness by aggregating multiple base results. However, traditional methods treat ensemble features statically, failing to account for the inconsistencies and variations that arise from differing base clustering outputs. To tackle this, we propose a novel iterative ensemble clustering framework that adaptively refines consensus representations via high-order graph diffusion and feedback-driven refinement. Unlike static consensus methods, our approach uniquely embeds evolving cluster structures within the feature space, enabling progressive self-correction through global inconsistency suppression and reliability-aware weighting. Our approach uniquely integrates: (1) Adaptive affinity weighting to prioritize reliable base clusterings, (2) Higher-order graph diffusion to capture topological dependencies and suppress high-frequency noise, and (3) Feedback-driven consensus evolution that iteratively replaces unreliable base clusterings with improved consensus outputs. Our method consistently outperforms 9 state-of-the-art ensemble techniques across 10 multimedia datasets, achieving \(\mathbf {+2.2-17.8\%}\) accuracy and \(\mathbf {+3-11.6\%}\) consensus quality, alongside cluster balance improvement up to \(\mathbf {+10.5\%}\). The proposed framework is scalable, model-agnostic, and ideal for real-world multimedia tasks with heterogeneous data sources.

External IDs:doi:10.1145/3743093.3771061