Augmenting Sampling with Link Prediction in Social Networks

• Activity Number: 391
• Type: Topic Contributed
• Date/Time: Tuesday, July 31, 2012 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): Tyler McCormick*+ and Ali Shojaie
• Companies: University of Washington and University of Washington
• Keywords: Hierarchical modeling ; Link prediction ; Sampling ; Social network

Abstract

Collecting complete social network data (an exhaustive list of links between actors) remains financially or logistically prohibitive in the majority of health and social science contexts. In response to this dilemma, scientists increasingly collect partially observed social network data through sampling. These methods typically produce small subgraphs, which consist of an exhaustive set of connections between a subset of nodes. Though the sampling mechanism is local, the goal of analysis remains global, understanding an influence process on a network or learning about a specific network feature (such as degree distribution or length of the longest path). We propose a novel statistical approach for the analysis and design of partially observed social network data. Our method first uses improved link prediction methods to probabilistically expand and connect subgraphs. We then use the expanded subgraphs to estimate network features. An overarching hierarchical model passes uncertainty between the two stages and improves estimation in sparse graphs. We demonstrate the effectiveness of our method by simulating sampling from several complete-network data sources.