Local Dependence in Large Event Data

• Applicant: Dr. Cornelius Fritz
• Subject Area: Statistics and Econometrics
• Term: from 2023 to 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 520770522

Data from online applications, websites, and sociometric badges enable the automatic collection of large event data containing valuable information about social interactions and behavior patterns. This data can be perceived as a dynamic network between a fixed number of actors, whose analysis using innovative and scalable methods is becoming increasingly important. In this context, events are understood as dyadic interactions between two actors. Current state-of-the-art techniques to study event data assume that every actor is aware of all other actors in the network and the events between them. This project attempts to close the resulting methodological gap. Since this assumption of globally available information is unreasonable for large event networks, we relax it by relying on local dependence. In this form of dependency, we assume that dependence between events is restricted to neighborhoods, leading to several advantages: The models are scalable in that they exhibit reasonable behavior even for large networks. Furthermore, the methods are scalable, allowing their estimation in the context of larger networks. Local dependencies allow for parallelized computation, which facilitates estimation with large data sets. Additionally, we can derive theoretical properties that enable inference on unseen data. We develop methods to utilize different forms of local dependencies for event data, which are distinguished by three types of neighborhoods. First, we assume that the neighborhoods are non-overlapping and latent. Second, we propose models with overlapping dependence structures that are determined by domain-driven mechanisms from each actor's point of view. An example of this mechanism could be that actors know the events of other actors if they have previously interacted or have common partners. Last, latent variable models represent hierarchically nested neighborhood structures. Overall, this project aids structural theory-building and hypothesis-testing with event data under realistic assumptions while providing practical tools, such as recommender systems to predict the next event and freely available software implementations.

• DFG Programme: WBP Fellowship
• International Connection: Switzerland, USA
• Hosts: Professor Dr. David Hunter; Professor Dr. Christoph Stadtfeld