Mechanistic mathematical models of opinion dynamics have been an important tool to gain insight into the qualitative dynamics of the evolution of opinions or ideologies over time. In order to study and understand the underlying mechanisms of the evolution of opinions, one needs interpretable models to be able to combine the insights from data with analysis. How can we select and validate reasonable mathematical models for a social system where opinions vary in time? How can we understand the interaction structure in these systems when information about the network is not available? In this talk, we consider the problem of inferring interaction kernels in an asynchronous agent-based model of pairwise opinion dynamics on a known network. Through synthetic experiments, we demonstrate that likelihood maximization is often able to accurately approximate the ground-truth kernel, even in cases where the inferred parameters do not match ground truth due to lack of identifiability. We employ an expectation-maximization algorithm to perform the likelihood-optimization step. We demonstrate our framework on simulation data generated by several kernels from well-studied models of opinion dynamics, including consensus dynamics, bounded-confidence dynamics, and attraction-repulsion dynamics.
Copyright. All Rights Reserved.