Predicting Propensity to Vote with Supervised Machine Learning

Abstract

This study explores the potential of supervised machine learning (ML) in predicting the propensity of a registered voter to vote by training over the IPUMS-ASA U.S. Voting Behaviors dataset. The investigation encompasses three distinct experiments, each varying in using the original dataset and derived data columns, including race, marital status, Hispanic origin, educational level, and high school graduation status. The initial experiment, involving no modifications to the original dataset, achieved an accuracy of 68.6% and produced the highest Matthews Correlation Coefficient (MCC) in the study, reflecting a moderately accurate correlation. Unexpectedly, experiments that attempted to enhance accuracy by adding contrived data reported reduced MCC scores. These findings underscore the potential of supervised ML in predicting voter behavior and offer a foundational framework for further research in refining supervised learning methodologies within political science. Future work might include integrating a broader array of datasets and experimenting with alternative model architectures to establish best practices across various contexts.

Keywords: GOTV, supervised machine learning, TensorFlow, elections, voter, prediction