Efficient Local Coherent Structure Learning via Self-Evolution Bipartite Graph

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Dimensionality reduction (DR) is crucial for machine learning tasks like classification and clustering.
• DR methods assuming Gaussian distributions struggle with Non-Gaussian data.
• Existing graph-based DR methods face challenges in finding optimal graphs and subspaces due to pairwise point comparisons, increasing complexity.

Purpose of the Study:

• To develop a novel dimensionality reduction approach that effectively handles Non-Gaussian data.
• To improve the efficiency and performance of DR methods by addressing limitations of current graph-based techniques.
• To introduce a new graph learning strategy that explores local data structures more effectively.

Main Methods:

• Proposed a novel self-evolution bipartite graph (SEBG) utilizing anchor points for landmark subclasses.
• Developed an efficient local coherent structure learning (ELCS) algorithm based on SEBG for automatic graph edge updates in learned subspaces.
• Employed a multivariable iterative optimization algorithm with theoretical proofs for problem-solving.

Main Results:

• The SEBG approach learns anchor-based relationships, improving locality exploration efficiency over pairwise methods.
• The ELCS algorithm demonstrates automatic updating of graph edges within the learned subspace.
• Extensive experiments show superior performance and efficiency compared to state-of-the-art methods on benchmark and large-scale image datasets.

Conclusions:

• The proposed SEBG and ELCS offer a significant advancement in dimensionality reduction for Non-Gaussian data.
• The anchor-based graph learning strategy enhances efficiency and effectiveness in exploring local data structures.
• The method achieves state-of-the-art results, proving its value for complex machine learning applications.