Mining Human Decision Making

• In Event: Deep Learning in Political Science

Abstract

We describe and test the components of human decision making using the text of an input and the human decision made from that text as an
outcome. To do this, we apply recurrent convolutional neural nets recently developed to model the meanings of non-consecutive groups of
words, and an intermediate layer specifically designed to make neural network classifications more interpretable. Specifically, we use a
network designed to extract the subset of a text that maximizes predictive information while minimizing the amount of text chosen. Instead
of using these algorithms to better understand classification, we use them to extract coherent phrases and sentences that are highly
predictive of a human decision. We show how clustering these phrases can describe the factors of human decision making in domains where
phrases or sections of text are pivotal in the decision. We apply our methods to replicate a conjoint experiment on immigration using
unstructured text, to decisions made in U.S. local bureaucracy, and to censorship decisions on Chinese social media. We discuss how they
could be use to generate research designs for experiments.

Authors

• Margaret E. Roberts, University of California, San Diego

• Luke Sanford, UC San Diego