Generalization performance of magnitude-preserving semi-supervised ranking with graph-based regularization

Semi-supervised ranking is a relatively new and important learning problem inspired by many applications. We propose a novel graph-based regularized algorithm which learns the ranking function in the semi-supervised learning framework. It can exploit geometry of the data while preserving the magnitude of the preferences. The least squares ranking loss is adopted and the optimal solution of our model has an explicit form. We establish error analysis of our proposed algorithm and demonstrate the relationship between predictive performance and intrinsic properties of the graph. The experiments on three datasets for recommendation task and two quantitative structure–activity relationship datasets show that our method is effective and comparable to some other state-of-the-art algorithms for ranking.