Electric network classifiers for semi-supervised learning on graphs

We propose a new classifier, named electric network classifiers, for semi-supervised learning on graphs. Our classifier is based on nonlinear electric network theory and classifies data set with respect to the sign of electric potential. Close relationships to C-SVM and graph kernel methods are revealed. Unlike other graph kernel methods, our classifier does not require heavy kernel computations but obtains the potential directly using efficient network flow algorithms. Furthermore, with flexibility of its formulation, our classifier can incorporate various edge characteristics; influence of edge direction, unsymmetric dependence and so on. Therefore, our classifier has the potential to tackle large complex real world problems. Experimental results show that the performance is fairly good compared with the diffusion kernel and other standard methods.