Explainable machine learning for predicting voting intentions: a study of Italian politics

Abstract Understanding voting intentions is a central challenge at the intersection of political science and machine learning. We introduce an explainable framework that integrates demographic variables with value-based attitudes to identify the ideological drivers of Italian electoral behaviour. Using 4500 survey responses collected between 2017 and 2019, we train ensemble classifiers with Bayesian optimisation and apply explainable AI methods. SHAP enhances interpretability at the individual level, while recursive feature elimination yields compact models. Results show that values concerning globalisation, immigration and populism improve predictive accuracy beyond demographics alone. Reduced feature sets retain performance and reveal coalition-specific voter profiles. We provide code and anonymised data to ensure reproducibility.