Learning From Small Samples: A Machine Learning Model of Belief Polarization

Kim, David

Etd

Learning From Small Samples: A Machine Learning Model of Belief Polarization

Pubblico Deposited

Why do people polarize? A common assumption is that people observe divisive information and form divisive beliefs. However, empirical studies indicate people can polarize even when they observe similar information. One possible reason is that people have cognitive biases and think irrationally. For instance, confirmation bias suggests prior beliefs influence how people interpret new evidence. We propose an alternative explanation for belief polarization. In an increasingly connected world, people are exposed to an abundance of information concerning a multitude of subjects. However, processing information is costly, so people may rely on small samples to form beliefs. Since small samples do not accurately represent the population and are variable, people may draw divergent images of the objective reality. First, we support our hypothesis using evidence from cognitive science literature. Then, we create a belief polarization model to test our hypothesis. We explore a unique approach and design a belief formation model based on machine learning. We propose new evaluation metrics for polarization and run simulations to observe how sample size affects polarization in various learning settings. Our results align with our hypothesis that, under basic assumptions, small sample reliance increases polarization. We offer practical suggestions for mitigating polarization based on our findings.

Creator