IDEAS SoCal Data Science Conference 2017

Understanding the mechanisms and drivers of human behavior is a difficult problem, accentuated by the heterogeneous nature of human behavioral data. This poses major issues for our ability to model and understand social systems, with important implications for design, testing and interventions in such systems.In this talk, I will present a statistical methodology to understand human behavior that quantifies feature importance, feature correlations and the level of predictability in human behavior data.

This methodology is highly interpretable, utilizing the R2 coefficient of determination to measure both the predictability of a system and the cumulative contribution of each feature towards this overall predictability. Our approach is non-parametric and free of any functional form, thus allowing for the capturing of non-linear and heterogeneous data which regularly occurs in human behavioral dynamics.

To illustrate, I will show our applications of this approach to various domains including the analysis of human performance in the Stack Exchange online forum and information sharing in the Twitter and Digg online social networks. In the case of information sharing on Twitter, for example, we show how our method effectively uncovers correlations among both individual-based features and information-based features, presenting a hierarchy of features that cumulatively explain human behavior in this social system.