p -Laplacian Regularization for Scene Recognition.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• The proliferation of internet multimedia data necessitates advanced machine learning algorithms, particularly for scenarios with scarce labeled samples.
• Manifold Regularized Semi-Supervised Learning (MRSSL) leverages data distribution structure to enhance model generalization.
• Existing MRSSL methods face challenges in fully exploiting the local geometry of data manifolds.

Purpose of the Study:

• To introduce an efficient approximation algorithm for graph p-Laplacian.
• To propose p-Laplacian regularization (pLapR) for preserving local data geometry in MRSSL.
• To enhance the generalization ability of machine learning models in low-label data scenarios.

Main Methods:

• Developed an efficient approximation algorithm for the graph p-Laplacian, reducing computational cost.
• Proposed p-Laplacian regularization (pLapR), a generalization of graph Laplacian, to better preserve local data structure.
• Applied pLapR within Support Vector Machines and Kernel Least Squares frameworks.

Main Results:

• The proposed pLapR method demonstrated superior preservation of local data geometry compared to conventional manifold regularization techniques.
• Experiments on Scene 67, Scene 15, and UC-Merced datasets confirmed the effectiveness of pLapR.
• The efficient approximation of graph p-Laplacian significantly reduced computational demands.

Conclusions:

• pLapR offers a theoretically sound and practically efficient approach to manifold regularized semi-supervised learning.
• The method shows significant improvements in machine learning tasks like scene recognition, especially with limited labeled data.
• pLapR advances the field of semi-supervised learning by providing a robust way to exploit data manifold structures.