This software learns the best of (finitely many) graphs for semi-supervised learning (within the framework of Belkin, Matveeva and Niyogi). These graphs may be constructed, for example, by applying a distance metric on points in Euclidean space and then computing k nearest neighbors or exponentially decaying weights.
Our methodology is based on multiple kernel learning and the connection between graph Laplacians and reproducing kernels. The algorithm is described in the paper Combining Graph Laplacians for Semi-Supervised Learning. The code also includes implementation of a few image transformations such as tangent distances.
To run the experiments from the paper execute usps_experiments
after downloading the USPS data set and creating the required mat files.