Label Information Guided Graph Construction for Semi-Supervised Learning.

Area of Science:

• Machine Learning
• Computer Science
• Data Science

Background:

• Existing graph-based semi-supervised learning methods often overlook label information during the graph learning phase.
• This oversight limits the potential for accurately capturing data structures and improving learning performance.

Purpose of the Study:

• To propose a novel semi-supervised graph learning method that integrates label information directly into the graph construction process.
• To enhance the performance of semi-supervised learning by better utilizing the inherent structure of the data.

Main Methods:

• Incorporated label information into graph learning by setting edge weights between labeled samples of different classes to zero.
• Adapted state-of-the-art methods like Low-Rank Representation (LRR) to include this label constraint.
• Formulated the problem as a convex optimization problem solvable via the linearized alternating direction method.

Main Results:

• The proposed semi-supervised low-rank representation method effectively captures the global geometric structure of data.
• Experimental results on synthetic and real datasets show superior performance in semi-supervised learning tasks compared to existing methods.
• Demonstrated the generalizability of the approach to various self-representation graph learning techniques.

Conclusions:

• Integrating label information during graph learning is a beneficial strategy for semi-supervised learning.
• The proposed method offers a robust and effective way to improve semi-supervised learning performance by enhancing graph structure representation.
• The approach is broadly applicable to existing self-representation graph learning frameworks.