Our approach to semi-supervised learning
Our approach uses some labeled nodes to propagate the information over the graph disambiguating unlabeled nodes in a consistent way. This process is based on two fundamental principles: the homophily principle, borrowed from social network analysis, and the transductive learning. The former simply states that similar objects are expected to have the same class [EAS 10]. We extended this principle assuming that objects, which are similar, are expected to have a similar class; an idea used also by [KLE 02], within a Markov random field framework. The latter is a case of semi- supervised learning [SAM 11] particularly suitable for relational data (see section 6.3.1).
In our system, we used a graph to model the geometry of the data and an evolutionary process to propagate the information over it. The graph construction method is described in section 6.4.1 and the evolutionary process in section 6.4.4. This work extends our previous works on unsupervised and semi-supervised WSD [TRI 15a, TRI 17].