Weighted ensemble clustering with multivariate randomness and random walk strategy

Abstract

Ensemble clustering algorithms have made significant progress in recent years due to their excellent performance. However, most of these algorithms face two challenges: one is to focus on the selection of subspaces since there is limited discussion on how to construct a potential metric space, the other is to treat basic clustering equally without fully considering the local connection between clusters when constructing the cooperative association matrix. To solve these issues, we propose a weighted ensemble clustering algorithm with multiple randomness and random walk strategy. We define the free exponential similarity kernel to create a diverse set of random metric spaces coupled with random subspaces and use spectral clustering to generate base clustering. Moreover, we use random walk strategy to discover the local connection between clusters and weight the collaborative association matrix. Finally, the collaborative association matrix uses consensus functions based on hierarchical clustering and meta clustering to obtain clustering results. On this basis, two specific ensemble clustering algorithms WECMR-HC and WECMR-MC are proposed. Theoretical analysis and experimental results demonstrate that our proposed algorithm outperforms existing ensemble algorithms.