What is semi-supervised learning on graphs? We have been told that if we just have a function which has an input graph, or a given graph with labeled nodes, we should be able to predict labels on other nodes of the graph. This is not clear.
Semi-supervised learning means that you get both labelled and unlabelled data as your input. The graph part means that there is also a relation defined for your datapoints. What exactly it is that you learn and how you use the graph structure depends on the problem. Here is a thorough exploration of the subject: http://www.cs.cmu.edu/~zhuxj/pub/thesis.pdf