Link Prediction Using Network Topology and Node Covariates

• Activity Number: 568
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Bopeng Li*+ and Ambuj Tewari
• Companies: and University of Michigan
• Keywords: link prediction ; node covariates ; support vector machine ; multiple kernel learning

Abstract

Link prediction is an important and open problem in statistical network analysis. Many methods have been proposed for solving the problem, but most of them only take into account measures computed using the network topology, such as the Jaccard distance, while paying little attention to additional information on the nodes that is often available in practice. We develop a multiple kernel learning (MKL) based method that uses both network topology and node covariates for link prediction. We show how these two types of covariates can be combined to achieve better prediction accuracy than using either type of covariates alone. We also propose an approach to address the class imbalance problem, since many real world networks have a much smaller number of edges than non-edges, by optimizing a loss function with different weights on different classes. We validate our method using experiment results on simulated and real network data.