TFMKC: Tuning-Free Multiple Kernel Clustering Coupled With Diverse Partition Fusion

Junpu Zhang,Liang Li,Pei Zhang,Yue Liu,Siwei Wang,Changbao Zhou,Xinwang Liu,En Zhu

DOI: https://doi.org/10.1109/TNNLS.2024.3435058

2024-08-23

Abstract:Clustering is a popular research pipeline in unsupervised learning to find potential groupings. As a representative paradigm in multiple kernel clustering (MKC), late fusion-based models learn a consistent partition across multiple base kernels. Despite their promising performance, a common concern is the limited representation capacity caused by the inflexible fusion mechanism. Concretely, the representations are constrained by truncated- k Eigen-decomposition (EVD) without fully exploiting potential information. An intuitive idea to alleviate this concern is to generate a set of augmented partitions and then select the optimal partition by fine-tuning. However, this is overlimited by: 1) introducing undesired hyperparameters and dataset-related consequences; 2) neglecting rich information across diverse partitions; and 3) expensive parameter-tuning costs. To address these problems, we propose transforming the challenging problem of directly determining the optimal partition (optimal parameter) into a diverse partition fusion (parameter ensemble) problem. We design a novel flexible fusion mechanism called tuning-free multiple kernel clustering coupled with diverse partition fusion (TFMKC) by reweighting diverse partitions through optimization, achieving an optimal consensus partition by integrating diverse and complementary information rather than traditional fine-tuning, and distinguishing our work from existing methods. Extensive experiments verify that TFMKC achieves competitive effectiveness and efficiency over comparison baselines. The code can be accessed at https://github.com/ZJP/TFMKC.