Pseudo-Supervision Affinity Propagation for Efficient and Scalable Multiview Clustering.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Anchor graph-based multiview clustering (AGMVC) offers efficiency but faces limitations in similarity measurement, scalability, and generalization.
• Existing methods struggle with single-structure information, leading to instability and high computational costs for large datasets.

Purpose of the Study:

• To develop an improved anchor graph construction method for multiview clustering that addresses limitations of existing approaches.
• To enhance the stability, efficiency, and generalization ability of multiview clustering algorithms.

Main Methods:

• Proposes a novel anchor graph construction learning local and global (LG) structures simultaneously.
• Introduces a landmark learning method for structural anchors, eliminating graph partitioning and the out-of-sample problem.
• Develops a pseudo-supervision affinity propagation (PSAP) framework to jointly optimize graph construction and landmark learning, accelerating convergence.

Main Results:

• The PSAP framework effectively disentangles in-cluster distributions between samples and anchors.
• A clustering inference partition (CIP) strategy is introduced for direct clustering output, avoiding postprocessing.
• Extensive experiments confirm the framework's efficiency and effectiveness in multiview clustering tasks.

Conclusions:

• The proposed LG structure learning and PSAP framework significantly advance anchor graph-based multiview clustering.
• The method offers a more stable, efficient, and generalizable solution for complex clustering problems.
• Publicly available code facilitates further research and application of the proposed framework.