Joint Dynamic Manifold and Discriminant Information Learning for Feature Extraction.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Neighborhood reconstruction is key for learning local manifold structures.
• Existing methods struggle with sparse coefficients, heterogeneous samples, and pre-dimensionality reduction manifold learning.

Purpose of the Study:

• To propose a joint model for simultaneous learning of affinity, reconstruction relationships, and projection matrices.
• To address limitations of traditional representation-based discriminant analysis for dimensionality reduction.

Main Methods:

• A joint model is developed to learn affinity and reconstruction relationships concurrently with the projection matrix.
• Active neighbor assignment and label information are used to learn inter-reconstruction coefficients.
• Sparse and whitening constraints are applied to reconstruction coefficients and the projection matrix, respectively.

Main Results:

• The proposed method effectively ensures local sparsity of reconstruction coefficients.
• It prevents non-zero coefficients between heterogeneous samples.
• Experiments demonstrate superior performance compared to existing methods on toy and public datasets.

Conclusions:

• The joint model offers a more accurate manifold structure by integrating dimensionality reduction.
• It enhances the performance of dimensionality reduction by addressing noise and redundant features.
• The proposed iterative algorithm provides a robust solution for this complex optimization problem.