Semi-supervised learning by edge domination in complex networks

Bio-inspired dynamical processes are able to identify nonlinear features in data. We present a dynamical process model of particle competition in complex networks applied to transductive semi-supervised learning. Particles carry labels and compete for the domination of edges. The process results consist of sets of edges arranged by label dominance. The sets are analyzed as subnetworks for the data classification. Computer simulations show that this model can identify nonlinear data forms in both real and artificial data, including overlapping structure of data.