Graph Neural Networks to Detect Suspicious Patterns in Nigerian Elections

The 2023 Nigerian Presidential Election generated verifiable and compelling allegations of polling unit-level result manipulation in many states in Nigeria. This study presents a Graph Neural Network (GNN) approach to detect that manipulation automatically because the central observation is that electoral suspicious patterns in Nigeria does not happen at random or isolated polling units. It happens in clusters, because ward collation officers can alter or replace multiple result sheets under their authority simultaneously. That coordinated, geographically clustered nature of suspicious patterns is exactly what GNNs are built to exploit. This study models each polling unit as a node in a graph where edges connect units that share the same ward. A 2-layer GATv2Conv model (available in Pytorch) then learns suspicious patterns patterns across ward neighbourhoods rather than individual units in isolation. Labels are engineered from three signals: statistical anomaly thresholds, Isolation Forest pre-screening, and critically, forensic annotations from the Centre for Collaborative Investigative Journalism (CCIJ) as gold labels. Training on Lagos (11,911 polling units) and Rivers (4,769 polling units) produces Area Under the Precision Recall Curve, AUC-PR of 0.9648 and 0.9196 respectively, representing 12.2x and 6.5x lifts over random baselines. An ablation study shows that adding ward membership edges alone accounts for a 132% improvement in AUC-PR over a feature-only MLP baseline. Ward-level suspicious patterns maps produced by the model independently converges on the specific wards named in CCIJ's forensic reporting, including Rumueme Ward 7A in Obio/Akpor at mean suspicious patterns probability of 0.998.