Oral Prelim: Bopeng Li, Combining Network Topology and Node Features for Link Prediction via Multiple Kernel Learning

In a real network (like Facebook), often both topological features (such as common neighbors) and node features (such as people's age, gender, interests, etc.) are available. These two sources of information are both useful in predicting links in a network. Therefore it is natural to combine them in order to get better prediction performance. We propose a link prediction method that combines network topology and node features in a principle and flexible way using Multiple Kernel Learning (MKL). Under this MKL framework, the optimal way of combining these two sources of information is well defined and can be learned from the training data using a modified version of the Generalized Multiple Kernel Learning (GMKL) algorithm. Experiments on both simulated and real networks (NIPS co-authorship network) show effectiveness of our method over baselines that either use only topology or only node features.