Scaling up graph-based semisupervised learning via prototype vector machines.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Semisupervised learning (SSL) leverages unlabeled data to enhance model performance when labeled data is scarce.
• Graph-based SSL methods, which enforce smoothness on data manifolds, achieve high accuracy but suffer from computational complexity.
• Scaling these methods to large datasets remains a significant challenge in machine learning.

Purpose of the Study:

• To develop a scalable graph-based semisupervised learning approach for large datasets.
• To reduce the time and space complexity of existing graph-based SSL techniques.
• To introduce a method for efficient data representation using sparse prototypes.

Main Methods:

• Approximating the data manifold as a weighted graph.
• Utilizing sparse prototypes as data representatives to approximate graph-based regularizers.
• Employing a principled method for prototype selection with Gaussian kernels.
• Controlling model complexity through the use of these prototypes.

Main Results:

• Achieved significant improvements in training and testing efficiency for semisupervised learning.
• Demonstrated encouraging performance and scalability on various real-world datasets.
• Showcased favorable comparisons with L1-regularization models at similar sparsity levels.
• Validated the efficacy of sparse prototypes in creating parsimonious and accurate SSL models.

Conclusions:

• Sparse prototypes offer an effective solution for scaling graph-based semisupervised learning.
• The proposed method enhances computational efficiency without compromising model accuracy.
• This approach facilitates the development of highly parsimonious and accurate models for semisupervised learning applications.