Comparing Diffusion Models for Graph–Based Semi–Supervised Learning

The main idea behind graph-based semi–supervised learning is to use pair–wise similarities between data instances to enhance classification accuracy (see (Zhu, 2005) for a survey of existing approaches). Many graph–based techniques use certain type of regularization that often involve a graph Laplacian operator (e.g., see (Belkin et al., 2006)). Intuitively, this corresponds to a diffusion process on graphs, where the information is propagated from the labeled instances to the rest of the nodes. Usually, this information is represented as a continuos class–membership probabilities (or scores), and the propagation process corresponds to the diffusion of those scores through the graph.