Anchor-Based Multiview Subspace Clustering With Anchor-wise and Class-wise Alignments.

Area of Science:

• Data Mining
• Machine Learning
• Computer Vision

Background:

• Multiview subspace clustering shows promise but struggles with large datasets due to high computational complexity.
• Existing anchor graph methods assume identical anchor-class structures across views, leading to alignment errors.
• This assumption overlooks variations in anchor ordering and class representation between views.

Purpose of the Study:

• To propose a novel anchor-based multiview subspace clustering method (AMCA^2) for improved performance on large-scale datasets.
• To address the limitations of heuristic anchor graph alignment in existing methods.
• To enhance the accuracy of anchorwise and classwise alignments in multiview clustering.

Main Methods:

• Developed the AMCA^2 method for simultaneous anchorwise and classwise alignment and fusion of multiple anchor graphs.
• Utilized permutation matrices and the Hadamard product for precise graph alignment.
• Introduced a novel kernel anchor selection (KAS) method for selecting more representative anchors.

Main Results:

• The proposed AMCA^2 method demonstrates superior performance compared to state-of-the-art techniques.
• Experiments on ten benchmark datasets validate the effectiveness of the new approach.
• The KAS method improves the selection of representative anchors, further boosting clustering accuracy.

Conclusions:

• AMCA^2 effectively overcomes the limitations of previous anchor graph strategies in multiview subspace clustering.
• The method achieves accurate anchorwise and classwise alignments, leading to improved clustering results.
• The proposed approach offers a more robust and scalable solution for large-scale multiview data analysis.