Transductive learning via spectral graph partitioning

We present a new method for transductive learning, which can be seen as a transductive version of the k nearest-neighbor classifier. Unlike for many other transductive learning methods, the training problem has a meaningful relaxation that can be solved globally optimally using spectral methods. We propose an algorithm that robustly achieves good generalization performance and that can be trained efficiently. A key advantage of the algorithm is that it does not require additional heuristics to avoid unbalanced splits. Furthermore, we show a connection to transductive Support Vector Machines, and that an effective Co-Training algorithm arises as a special case. 1