Applying Machine Learning to Election Forensics Research: A Case of Russia

• In Event: Autocratic Challenges in the Post Soviet States

Abstract

This research investigates the prediction accuracy of different election forensics methods designed to detect and measure the occurrence of election fraud, placing particular emphasis on the Bayesian version of the finite mixture model developed by Walter Mebane and his students, as well as the nonparametric approach based on vote-turnout histograms proposed by Sergey Shpilkin. Utilizing data from Russian election observers, the study employs diverse machine learning algorithms (such as boosted decision trees, neural networks, support vector machines, etc.), offering varied perspectives on the classification problems associated with the measurement of election fraud using various election forensics methods. This research serves to bridge the gap between the realms of election forensics and machine learning, offering insights into the detection of irregularities within electoral processes. The outcomes of this study are poised not only to make a significant contribution to academic discourse but also to have practical implications, fostering the development of robust tools essential for safeguarding the integrity of elections on a global scale.

Author

• Kirill Kalinin, Stanford University