Dirichlet mixtures of graph diffusions for semi supervised learning

Graph representations of data have emerged as powerful tools in the classification of partially labeled data. We give a new algorithm for graph based semi supervised learning which is based on a probabilistic model of the process which assigns labels to vertices. The main novelty is a non parametric mixture of graph diffusions, which we combine with a Markov random field potential. Markov chain Monte Carlo is used for the inference, which we demonstrate to be significantly better in terms of predictive power than the maximum a posteriori estimate. Experiments on bench-mark data demonstrate that while computationally expensive our approach can provide significantly improved predictions in comparison with previous approaches.