# Anomaly Detection and Automated Labeling for Voter Registration File Changes

**Authors:** Sam Royston, Ben Greenberg, Omeed Tavasoli, Courtenay Cotton  
**arXiv:** [2106.15285](https://arxiv.org/abs/2106.15285)  
**Date:** 2021-06-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.15285v1/

## Key points

- Voter eligibility process in the United States is managed through state databases
- Accuracy and security of these databases is challenging for administrators
- Detecting and monitoring improper modifications to Voter Registration Files (VRFs) is crucial
- Machine learning techniques can assist in protecting voter rolls
- Methods involve comparing snapshots of VRFs over time and using unsupervised anomaly detection models
- Statistical models and non-negative matrix factorization are effective in surfacing anomalous events
- Successful deployment during 2019-2020 in collaboration with the office of the Iowa Secretary of State
- Newly deployed model incorporates historical and demographic metadata to label root cause of database modifications
- Classifier trained to predict event labels for voter deactivations with high accuracy and F1-score (0.892)
- Advancements in machine learning-based approaches improve accuracy, security, and monitoring of voter registration databases

## Summary

The voter eligibility process in the United States is managed through state databases that contain information about eligible citizens. Ensuring the accuracy and security of these databases is a challenging task for administrators at the state and local level. Detecting and monitoring any improper modifications to the Voter Registration Files (VRFs) is crucial, as malicious actors could manipulate these files to disrupt the democratic process. While election officials performed admirably during the contentious 2020 elections, there is still much work to be done in securing and monitoring the election systems that Americans rely on. To address this challenge, a set of methods that utilize machine learning techniques have been developed to assist analysts and administrators in protecting voter rolls. These methods involve comparing snapshots of VRFs over time and using unsupervised anomaly detection models to identify anomalous changes. Statistical models that compare administrative districts within a short time span and non-negative matrix factorization have proven to be effective in surfacing anomalous events for review. These methods were successfully deployed during 2019-2020 in an organization's monitoring system, in collaboration with the office of the Iowa Secretary of State. In addition to anomaly detection, a newly deployed model has been proposed that incorporates historical and demographic metadata to label the likely root cause of database modifications. This model aims to predict which modifications have known causes, enabling better identification of potentially anomalous modifications. Furthermore, a classifier was trained to predict event labels for voter features related to deactivations. The classification results showed an accuracy and F1-score of 0.892. Overall, these advancements in machine learning-based approaches provide valuable tools for improving the accuracy, security, and monitoring of voter registration databases. By leveraging data analysis techniques and incorporating historical metadata, administrators can better protect against potential threats to the integrity of US elections.