Transferable Linear Discriminant Analysis.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Linear Discriminant Analysis (LDA) is a common feature extraction technique.
• Traditional LDA struggles with data from different domains due to distribution discrepancies.
• Unlabeled data is not utilized in standard LDA, limiting performance improvements.

Purpose of the Study:

• To propose a novel Transferable Linear Discriminant Analysis (TLDA) method.
• To extend LDA for scenarios with varying data probability distributions.
• To improve feature extraction and classification performance on cross-domain data.

Main Methods:

• TLDA utilizes a low-rank structure assumption for data within the same subspace.
• The method employs matrix rank as a criterion for local and global linear transformations.
• A projected subgradient-based optimization method is used for TLDA objective function.

Main Results:

• TLDA effectively reduces distribution discrepancies within subspaces, aligning data.
• The method enlarges distances between different subspaces, achieving maximal separation.
• Experimental results show TLDA outperforms state-of-the-art methods on public datasets.

Conclusions:

• TLDA offers a robust solution for feature extraction in cross-domain learning.
• The proposed method enhances classification accuracy by addressing distribution shifts.
• TLDA demonstrates superior performance compared to existing LDA techniques.