Self-paced and structured graph-based ensemble clustering

Ensemble clustering is an important topic in data mining, which combines multiple base clusterings to produce a more robust and effective clustering model. Scholars have proposed many methods in ensemble clustering and made important theoretical contributions in the past few years. However, there are still several drawbacks in these methods. Firstly, existing methods treat each base clusterings equally. Secondly, difficult instances are not fully considered, which may influence the optimization process. To address the aforementioned limitations, we introduce a novel self-paced and structured graph-based ensemble clustering algorithm (S²GEC), which introduces self-paced learning to gradually involve instances into training procedure from easy to difficult. In addition, we also proposed a unified optimization framework which can evaluate instances and base clusterings simultaneously so as to obtain the consensus clustering result. Moreover, S²GEC can obtain a structured similarity matrix to extract the clustering indicators directly. Extensive experiments conducted on commonly used datasets have demonstrated the superiority of our algorithm.