Propagation Kernels for Partially Labeled Graphs

Learning from complex data is becoming increasingly important, and graph kernels have recently evolved into a rapidly developing branch of learning on structured data. However, previously proposed kernels rely on having discrete node label information. Propagation kernels leverage the power of continuous node label distributions as graph features and hence, enhance traditional graph kernels to efficiently handle partially labeled graphs in a principled manner. Utilizing localitysensitive hashing, propagation kernels are able to outperform state-of-the-art graph kernels in terms of runtime without loss in prediction accuracy. This paper investigates the power of propagation kernels to classify partially labeled images and to tackle the challenging problem of retrieving similar object views in robotic grasping.