Consensus Affinity Graph Learning for Multiple Kernel Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Multiple Kernel Graph-Based Clustering (MKGC) combines Multiple Kernel Learning (MKL) and graph-based clustering.
• Existing MKGC methods often use complex 'fat' models, leading to high computational costs and suboptimal performance.
• There is a need for efficient and effective MKGC methods that directly learn a consensus affinity graph.

Purpose of the Study:

• To propose a new MKGC method that directly learns a consensus affinity graph, overcoming the limitations of existing 'fat' models.
• To preserve the local manifold structure of data in kernel space.
• To improve clustering performance and computational efficiency.

Main Methods:

• The proposed method first learns multiple candidate affinity graphs by preserving local manifold structure using self-expressiveness and adaptive local structure learning.
• A 'thin' autoweighted fusion model synthesizes these candidate graphs into a consensus affinity graph.
• Key techniques include a self-tuned Laplacian rank constraint and a top-k neighbors sparse strategy to enhance consensus graph quality.

Main Results:

• The proposed MKGC method was evaluated on ten benchmark and two synthetic datasets.
• Experimental results demonstrate consistent and significant performance improvements over state-of-the-art methods.
• The method effectively preserves local manifold structures and generates high-quality consensus affinity graphs.

Conclusions:

• The developed MKGC method offers a more efficient and effective approach to graph-based clustering.
• Directly learning a consensus affinity graph via a 'thin' model is advantageous for computational cost and clustering accuracy.
• The proposed techniques significantly advance the field of multiple kernel graph-based clustering.