Predicting Trust in Public Institutions in Eswatini Using Machine Learning and Deep Learning: A Comparative Analysis

Objectives: This study conducts a comprehensive analysis of artificial intelligence AI techniques for predicting and identifying factors impacting trust in public institutions in Eswatini (Swaziland). Although declining institutional trust has been widely documented across sub-Saharan Africa, empirical evidence for Eswatini remains limited, despite growing concerns related to governance, corruption, the COVID19 pandemic, and recent civil unrest. Methods: Utilizing Afrobarometer survey data from 2013 to 2022, Feature Tokenizer Transformer (FT-Transformer) and XGBoost were employed to predict institutional trust levels and to identify the most influential predictors using Information Value (IV), Attention Weights and permutation importance. Results: Results indicate that government performance indicators, economic perceptions, and corruption perceptions are significant determinants of trust. Comparative model evaluation shows that both models achieved competent performance (Accuracies: 0.68-0.76; AUC: 0.74-0.82), with XGBoost proving more computationally efficient and the FT-Transformer achieving higher peak performance for specific institutions like the Army. A significant shift in trust dynamics was observed for most institutions post-pandemic, though Local Government trust remained stable. Conclusion: The findings underscore the potential of AI-driven methods to enhance understanding of citizen trust dynamics and support data-informed governance in Eswatini, particularly in the aftermath of the COVID-19 pandemic