Network Inference from Grouping Data

• Activity Number: 237
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Government Statistics Section
• Author(s): Charles Weko*+
• Keywords: Network Inference ; Maximum Likelihood ; EM Algorithm ; Social Network

Abstract

In the past two decades the interest in network analysis has expanded rapidly. One particular area of research involves the inference of network structure from grouping data. Grouping data records the manner in which a population forms subsets or smaller groups. One important aspect of grouping data is that it frequently includes large groups that could have been produced in many different ways.

Typically, network inference is performed on grouping data through the use of descriptive statistics. Researchers have defined a collection of measures to calculate the frequency with which members of the population interact and use these measures to infer a network structure. Classic examples of these measures include the co-occurrence matrix and the half weight index.

In this work we define two stochastic models for group formation, the known star model and the unknown star model, and derive maximum likelihood estimators for their parameters.

Using a combination of simulated data and real world data, we solve for the parameters of the models using the Expectation-Maximization algorithm and show how the results differ from the co-occurrence matrix and the half weight index.