MetaWCE: Learning to Weight for Weighted Cluster Ensemble.

Cluster ensemble (CE) integrates multiple clustering solutions to effectively improve the accuracy and robustness of unsupervised clustering. To reduce the impacts of low-quality solutions, existing CE methods often design heuristic criteria to appraise these clustering solutions and allocate weights for them. However, such heuristic-based weighting methods rely on human experience and lack knowledge of the relation between weights and data characteristics, failing to adaptively adjust weights for various datasets. To address this, we propose Meta -learning-based W eighted C luster E nsemble (MetaWCE), which learns the weights-data relation automatically and sets adaptive CE weights. Specifically, metadata is employed to describe data characteristics at a dataset level. To bridge metadata and weights, a meta-learning strategy is introduced to simulate the weighting process to ensure that relation between weights and metadata can be learned to directly optimize the ensemble performance in an end-to-end manner. Experiments on three datasets indicate that MetaWCE significantly improves ensemble performance and achieves obvious improvements over strong baseline methods. • First meta-learning-based weighted cluster ensemble method. • Considers data characteristics to improve cluster ensemble performance. • Learns a domain-general metric space to extract clustering-friendly features. • Introduces a two-level meta weighting model to adaptively predict weights. [ABSTRACT FROM AUTHOR]