Semi-supervised bilinear subspace learning.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Tensor-based subspace learning methods like 2-D PCA and 2-D LDA have shown success in unsupervised and supervised learning.
• Existing semi-supervised algorithms often rely on regularization in the original feature space.

Discussion:

• This work proposes a new semi-supervised subspace learning algorithm by integrating tensor representation with unlabeled data.
• The algorithm utilizes graph Laplacian regularization in the low-dimensional feature space, differing from conventional approaches.
• An iterative algorithm, adaptive regularization based semi-supervised discriminant analysis with tensor representation (ARSDA/T), is developed for computation.

Key Insights:

• ARSDA/T effectively leverages complementary information from unlabeled data.
• The use of low-dimensional feature space for graph Laplacian regularization enhances performance.
• A vector-based variant (ARSDA/V) is also presented for tensor data conversion.

Outlook:

• ARSDA/T demonstrates significant improvements in face recognition accuracy on benchmark datasets (CMU PIE, YALE-B).
• The method offers a promising advancement over traditional supervised and semi-supervised subspace learning techniques.
• Future work may explore further applications of this tensor-based semi-supervised approach.