A Machine learning framework for electoral anomaly detection: case study using Israeli data for Kenyan electoral applications

Abstract

This dissertation investigates how machine learning algorithms can be harnessed to strengthen electoral integrity by proactively detecting anomalies in voting data. Motivated by recurrent concerns about fairness and transparency in Kenya’s electoral processes, the study applies a data driven approach to uncover patterns indicative of irregularities such as mismatches between registered voters and votes cast, suspicious voting turnouts, and inconsistencies in valid and invalid ballots. To model and test the detection framework, historical electoral data from Israel (1996– 2015) was used as a proxy due to its completeness and availability. The research followed the CRISP-DM methodology, encompassing phases of data understanding, preprocessing, algorithm training, and system deployment. The Isolation Forest algorithm, known for its unsupervised anomaly detection capabilities, was selected and adapted for the electoral context. The model successfully flagged 9,856 data points as anomalous across various election cycles, validating its applicability. To enhance usability, the algorithm was integrated into a Stream lit web application designed for interactive analysis, visualization, and stakeholder engagement. Through this deployment, electoral practitioners can upload datasets, visualize irregularities, and download reports in real time. The study contributes to the growing body of research on AI in public governance by presenting a practical, replicable model for anomaly detection in elections. It also proposes governance policy recommendations for adopting such tools in the Kenyan context, with the ultimate goal of fostering fairer, data-informed electoral oversight.