Semi-Supervised Gaussian Process Classifiers

In this paper, we propose a graph-based construction of semi-supervised Gaussian process classifiers. Our method is based on recently proposed techniques for incorporating the geometric properties of unlabeled data within globally defined kernel functions. The full machinery for standard supervised Gaussian process inference is brought to bear on the problem of learning from labeled and unlabeled data. This approach provides a natural probabilistic extension to unseen test examples. We employ Expectation Propagation procedures for evidence-based model selection. In the presence of few labeled examples, this approach is found to significantly outperform cross-validation techniques. We present empirical results demonstrating the strengths of our approach.