Interaction event times observed on a social network provide valuable information for social scientists to gain insight into complex social dynamics that are challenging to understand. However, it can be difficult to accurately represent the heterogeneity in the data and to model the dependence structure in the network system. This requires flexible models that can capture the complicated dynamics and complex patterns. Point process models offer an elegant framework for modeling event time data. This dissertation concentrates on developing point process models and related diagnostic tools, with a real data application involving an animal behavior network. In this dissertation, we first propose a Markov-modulated Hawkes process (MMHP) model to capture the sporadic and bursty patterns often observed in event time data. A Bayesian inference procedure is developed to evaluate the likelihood by using a variational approximation and the forward-backward algorithm. The validity of the proposed model and associated estimation algorithms is demonstrated using synthetic data and the animal behavior data.

Facilitated by the power of the MMHP model, we construct network point process models that can capture a social hierarchy structure by embedding nodes in a latent space that can represent the underlying social ranks. Our model provides a ranking method for social hierarchy studies and describes the dynamics of social hierarchy formation from a novel perspective – taking advantage of the detailed information available in event time data. We show that the network point process models appropriately captures the temporal dynamics and heterogeneity in the network event time data, by providing meaningful inferred rankings and by calibrating the accuracy of predictions with relevant measures of uncertainty. In addition to developing a sensible and flexible model for network event time data, the last part of this dissertation provides essential tools for diagnosing lack of fit issues for such models. We develop a systematic set of diagnostic tools and visualizations for point process models fitted to data in the dynamic network setting.

By inspecting the structure of the residual process and Pearson residual on the network, we can validate whether a model adequately captures the temporal and network dependence structures in the observed data.