Semi-supervised Penalized Output Kernel Regression for Link Prediction

Link prediction is addressed as an output kernel learning task through semi-supervised Output Kernel Regression. Working in the framework of RKHS theory with vectorvalued functions, we establish a new representer theorem devoted to semi-supervised least square regression. We then apply it to get a new model (POKR: Penalized Output Kernel Regression) and show its relevance using numerical experiments on artificial networks and two real applications using a very low percentage of labeled data in a transductive setting. 1.