Multiple Kernel Clustering With Compressed Subspace Alignment.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Multiple kernel clustering (MKC) effectively fuses multisource information for improved clustering.
• Existing MKC methods face scalability challenges due to high memory (O(n^2)) and computational (O(n^3)) complexity.
• This limits their application in median- and large-scale datasets.

Purpose of the Study:

• To redesign subspace segmentation-based MKC for improved efficiency and scalability.
• To develop an MKC algorithm with reduced memory (O(n)) and computational (O(n^2)) complexity.
• To enhance clustering performance and speed through a novel, integrated sampling strategy.

Main Methods:

• Redesigned the formulation of subspace segmentation-based MKC to achieve O(n) memory and O(n^2) computational complexity.
• Introduced a novel sampling strategy, mathematically modeled and learned simultaneously during information fusion.
• Leveraged GPU acceleration and multicore techniques for parallelization.

Main Results:

• The proposed algorithm demonstrates superior performance compared to state-of-the-art methods on six datasets.
• Achieved comparable time costs to linear complexity algorithms despite the integrated sampling process.
• The novel sampling strategy improved data reconstruction and discriminative capability across views.

Conclusions:

• The redesigned MKC algorithm offers a scalable and efficient solution for multisource information fusion in clustering.
• The integrated sampling strategy enhances both performance and speed, making MKC applicable to larger datasets.
• Parallelization capabilities further boost efficiency, positioning the algorithm as a competitive alternative for real-world applications.