Ensemble Clustering Based on Manifold Broad Learning System

Abstract

Ensemble clustering has made significant progress in improving clustering performance by integrating multiple basic partitions. However, the conventional ensemble clustering methods rarely combine the rich information from both basic partitions and feature space comprehensively and efficiently. Facing the limitation, we design an Ensemble Clustering Based on Manifold Broad Learning System (EC-MBLS) in this paper. EC-MBLS firstly constructs a fuzzy partition matrix to capture the cluster-wise similarities from basic partitions. Then a manifold broad learning system (MBLS) is designed to learn a better consensus partition, which is guided by the global cluster structure of data and constrained by the local affinity of data. An efficient solution is finally provided to optimize the MBLS. Compared with the state-of-the-art ensemble clustering methods, experiments on multiple real-world clustering tasks are carried out to demonstrate the effectiveness of EC-MBLS.