Multi-View Dynamic Kernelized Evidential Clustering

Abstract

It is challenging to cluster multi-view data in which the clusters have overlapping areas. Existing multi-view clustering methods often misclassify the indistinguishable objects in overlapping areas by forcing them into single clusters, increasing clustering errors. Our solution, the multi-view dynamic kernelized evidential clustering method (MvDKE), addresses this by assigning these objects to meta-clusters, a union of several related singleton clusters, effectively capturing the local imprecision in overlapping areas. MvDKE offers two main advantages: firstly, it significantly reduces computational complexity through a dynamic framework for evidential clustering, and secondly, it adeptly handles non-spherical data using kernel techniques within its objective function. Experiments on various datasets confirm MvDKE's superior ability to accurately characterize the local imprecision in multi-view non-spherical data, achieving better efficiency and outperforming existing methods in overall performance.