Multiple Kernel Clustering with Dual Noise Minimization

What is it about?

This paper investigated the dual noise within kernel partition in multiple-kernel clustering. We mathematically disassemble the noise into N-noise and C-noise, distinguishing our work from existing researches. And we proposed an elegant, compact, hyperparameter-free, and efficient algorithm to minimize the dual noise.

Featured Image

Photo by Jakob Owens on Unsplash

Photo by Jakob Owens on Unsplash

Why is it important?

For the first time, we mathematically disassemble the noise into N-noise and C-noise. We find that C-noise exhibits stronger destruction than N-noise on the block diagonal structures, which directly leads to the degeneration of clustering performance. To depress the impact of dual noise, we propose a novel elegant model to minimize dual noise in late fusion kernel framework. We propose an efficient two-step alternative optimization strategy to solve our model with linear computation complexity, and achieve state-of-art clustering performance on various benchmark datasets.