Conditional Random Field Models as a Way to Capture Peer Influence in Social Networks

Peer influence occurs when individuals adapt their behavior according to the behavior of their friends. When studying human interactions and the spreading of beliefs, feelings or behaviors, influence mechanisms typically play a decisive role. If there are multiple observations of network and behavior, temporal models such as SIENA can identify social influence based on behavioral change at different time points. However, in cross-sectional cases, where only one observation of the network is available, studying and predicting individual behavior while controlling for social influence is very challenging both statistically and computationally. In this work, we propose using Conditional Random Field (CRF) logistic regression for modeling peer influence in cross-sectional settings and compare it with existing methods such as Autologistic Actor Attribute Models (ALAAM).<br> We use data about teenage smoking behavior from previous social influence studies to evaluate our approach. The results of CRF models are consistent with ALAAM. CRF produces accurate coefficient estimations compared to the over-estimation in ordinary logistic regression models. For example, after controlling for contagion effects, gender has no significant impact on smoking; but it has in the ordinary logistics regression. Similarly, drinking alcohol, smoking siblings, and being in a romantic relationship have smaller effect sizes when social influence effects are taken into account. This study shows that CRF models are capable of modeling individual behavior with peer influence and are both computationally efficient and scalable for large networks. Moreover, the extension of the CRF models can characterize different types of peer influence such as contagion and similarity and can model dynamic behavior in a network.