Local Sample-Weighted Multiple Kernel Clustering With Consensus Discriminative Graph.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Multiple Kernel Clustering (MKC) aims to fuse information from base kernels effectively.
• Precise and local kernel matrices are crucial for clustering performance, as distant similarity estimations can be unreliable.
• Existing localized MKC methods often use the K-Nearest Neighbors (KNN) mechanism, which treats all neighbors equally, limiting performance.

Purpose of the Study:

• To propose a novel local sample-weighted MKC (LSWMKC) model.
• To address the limitations of existing KNN-based localization in MKC algorithms.
• To improve the accuracy and effectiveness of kernel-based clustering through adaptive neighbor weighting.

Main Methods:

• Constructing a consensus discriminative affinity graph in kernel space to reveal latent local structures.
• Outputting an optimal neighborhood kernel with sparse properties and a block diagonal structure.
• Implicitly optimizing adaptive weights for different neighbors based on corresponding samples.

Main Results:

• The proposed LSWMKC model demonstrates superior local manifold representation.
• LSWMKC outperforms existing kernel and graph-based clustering algorithms in experimental evaluations.
• The method effectively addresses the issue of equal neighbor importance in KNN-based localization.

Conclusions:

• LSWMKC offers a more refined approach to kernel matrix localization by adaptively weighting neighbors.
• The model achieves better clustering performance by capturing local structures more accurately.
• The proposed method advances the field of multiple kernel clustering and provides a valuable tool for data analysis.