Graph-Based Semi-Supervised Learning

In this chapter, we continue our discussion about semi-supervised learning, considering a family of algorithms that are based on the graph obtained from a dataset, and the existing relationships among samples. The problems that we are going to discuss belong to two main categories: the propagation of class labels to unlabeled samples, and the use of non-linear techniques based on the manifold assumption to reduce the dimensionality of the original dataset. In particular, this chapter covers the following propagation algorithms:

• Label propagation based on the weight matrix
• Label propagation in scikit-learn, based on transition probabilities
• Label spreading
• Laplacian regularization
• Propagation based on Markov random walks

For the manifold learning section, we're discussing the following:

• The Isomap algorithm and the multidimensional scaling approach
• Locally linear embedding
• Laplacian spectral embedding...