General subspace learning with corrupted training data via graph embedding.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Subspace learning aims to identify underlying data structures.
• Real-world data often contains corruptions, hindering traditional methods.
• Existing discriminant analysis algorithms struggle with significant data corruption.

Purpose of the Study:

• To develop a robust subspace learning method capable of handling corrupted training data.
• To introduce a novel supervised subspace learning algorithm, Corruptions Tolerant Discriminant Analysis (CTDA).
• To demonstrate CTDA's superiority over conventional methods in the presence of data corruptions.

Main Methods:

• Formulated subspace learning as a nuclear norm regularized optimization problem.
• Decomposed the learned subspace into intrinsic, penalty, and error components.
• Proposed CTDA, a supervised learning algorithm utilizing the learned subspace components.

Main Results:

• The proposed optimization problem is convex and solvable in polynomial time.
• CTDA effectively models intrinsic features, undesired properties, and data corruptions.
• Experiments show CTDA outperforms related algorithms on face and object recognition datasets.

Conclusions:

• CTDA offers a robust solution for supervised subspace learning with corrupted data.
• The method's ability to handle gross corruptions addresses limitations of prior algorithms.
• CTDA demonstrates strong performance in practical applications like image recognition.