Joint Projection Learning and Tensor Decomposition-Based Incomplete Multiview Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Incomplete multiview clustering (IMVC) is challenging due to missing data across views.
• Existing IMVC methods often suffer from suboptimal graph construction on high-dimensional data, leading to noise and feature redundancy.
• Previous methods overlook graph noise arising from structural variations during graph transformation.

Purpose of the Study:

• To propose a novel joint projection learning and tensor decomposition (JPLTD) method for robust IMVC.
• To address feature redundancy and noise in high-dimensional data for improved clustering.
• To mitigate graph noise caused by missing samples and structural variations.

Main Methods:

• Introduces an orthogonal projection matrix for compact feature learning in a lower-dimensional space.
• Learns similarity graphs from projected features and forms a third-order low-rank tensor to capture cross-view correlations.
• Employs tensor decomposition-based graph filtering to handle noise in projected data and identify intrinsic data similarities.

Main Results:

• The proposed JPLTD method effectively reduces the impact of redundant features and noise.
• Robust graph filtering enhances clustering performance by modeling true data similarities.
• Experiments on benchmark datasets show JPLTD outperforms existing state-of-the-art IMVC methods.

Conclusions:

• JPLTD offers a robust and effective solution for incomplete multiview clustering.
• The joint projection learning and tensor decomposition approach significantly improves clustering accuracy.
• The method's ability to handle feature redundancy and graph noise makes it suitable for real-world incomplete data scenarios.