Multiscale Laplacian Learning.

Area of Science:

• Machine learning
• Data science
• Computational science

Background:

• Machine learning (ML) excels in many fields but struggles with limited labeled data.
• Diverse and small datasets, common in high-cost or ethically constrained research, hinder ML performance.
• Existing ML methods often require large labeled datasets, posing a significant challenge.

Purpose of the Study:

• To develop innovative strategies for machine learning with limited, diverse, and small datasets.
• To introduce two novel multiscale Laplacian learning (MLL) approaches.
• To enhance data classification and tackle data scarcity challenges in ML.

Main Methods:

• Integration of graph-based frameworks, semi-supervised techniques, and multiscale structures.
• Development of multikernel manifold learning (MML) using multiscale graph Laplacians and a warped kernel regularizer.
• Adaptation of the Merriman-Bence-Osher (MBO) scheme with multiscale Laplacians and fast solvers (MMBO).

Main Results:

• Two novel MLL approaches, MML and MMBO, were developed.
• Experimental validation on benchmark datasets demonstrated the effectiveness of the proposed algorithms.
• The new methods achieved favorable comparisons against state-of-the-art approaches.

Conclusions:

• The proposed MLL methods offer effective solutions for machine learning tasks involving limited, diverse, and small datasets.
• MML and MMBO represent significant advancements in handling data constraints in machine learning.
• These approaches have broad applicability in scientific fields facing data limitations.