Agent-Based Probabilistic Models of Social Interaction

Multi-agent simulation is a powerful tool for studying real-world interactions and identifying influential factors that determine the emergent dynamics of social systems. While they are often partially validated using theories from the social sciences or correspondences with real-world data, it is usually difficult to accurately characterize real-world population phenomena with these models. This can be limiting in applications involving large, distributed systems, for example, if we want to know how the opinions of users are evolving over time in a social network. Social media provides an expansive and readily available dataset for analyzing such models, but very rarely are there compatibilities with the textual information available on social media and the latent characteristics that simulated agents exhibit. To develop more effective agent-based approach to modeling interaction and expression patterns on social media, we propose a probabilistic analysis framework in which the internal state of the agent and some contextual situation influences the textual content of their post. We investigate several text-based models that can be validated by and used to analyze social media corpora.