Discovering Connections Between Networks and Outcomes: Labeled Subgraph Kernels for Social Network Analysis

• In Event: 30802 - Machine Learning and Classical Statistical Approaches: Trade-offs, Integration, and Debate (Co-sponsored by Section on Mathematical Sociology)

Abstract

Application of classical multivariate analysis to network data has a long tradition within the social network community, with examples ranging from distance or similarity-based clustering or scaling of networks to linear network regression. These approaches are powerful tools for network comparison, particularly in exploratory settings and/or where large numbers of networks are involved. Many of these ideas can be unified within the framework of kernel learning, in which more general families of similarity measures (kernels) are used to extend linear procedures to broader classes of function spaces. Here, we discuss a particular family of kernels whose feature space consists of indicators for the presence of labeled subgraphs, and that is hence well-suited for many social network applications. We demonstrate the use of these kernels on concept networks derived from social media posts, showing how they can be used to integrate network data with other types of predictors in regression and classification settings.

Author

• Carter T. Butts, University of California-Irvine