Interpretable machine learning with SHAP predicts honest behavior from personality traits and physiological data

Abstract This paper investigates the extent to which dishonest behavior can be predicted using behavioral personality traits and physiological information. An interpretable machine learning approach, SHapley Additive exPlanations (SHAP) analysis, is employed to identify the most influential predictors. Drawing on data from the HEXACO personality inventory and physiological measures (including skin conductance activity and heart rate variability), we construct several machine learning models: a personality-based model, a physiology-based model, and a combined model. The results indicate that the AdaBoost personality-based model achieved the strongest predictive performance (AUC = 0.836), outperforming models based solely on physiological indicators. SHAP analysis further reveals that Openness to Experience and Extraversion are the most salient positive predictors of dishonest behavior. Among the physiological features, very-low-frequency components of heart rate variability and finger pulse amplitude emerge as meaningful predictors. Overall, this study offers preliminary insights into predictors of immediate dishonest behavior and suggests that integrating personality traits with physiological signals is a feasible approach for predictive modeling, providing a methodological reference for assessing individual honesty.