This study addresses the challenge of predicting cooperative behaviour in repeated strategic interactions by combining theory-driven and data-driven approaches. Cooperation, a central problem in behavioural science, is commonly studied using the Iterated Prisoner’s Dilemma (IPD), where individuals repeatedly choose between collective and self-interested actions. Traditional cognitive models, such as the functional Experience-Weighted Attraction (fEWA) model, provide interpretable insights into decision-making processes but often lack predictive accuracy. In contrast, machine learning models achieve strong predictive performance but offer limited interpretability.

To bridge this gap, this study develops a hybrid modelling framework that integrates fEWA-derived features with machine learning techniques. Using experimental IPD data, multiple models were trained on behavioural, contextual, and theory-informed features. Four model categories were compared: a basic machine learning model, an extended model with hand-crafted features, a simple hybrid model incorporating fEWA features, and a full hybrid model combining all feature types.

Results show that hybrid models outperform both standalone cognitive and basic machine learning approaches, with the full hybrid model achieving the highest predictive accuracy. However, improvements over the best purely data-driven model were modest, suggesting that well-designed behavioural features can capture much of the predictive signal provided by cognitive theory. Model interpretation using SHAP values further reveals the relative contribution of different feature types.

Overall, this study suggests that the inclusion of cognitive model features offers potential benefits for interpretability and theory development. These findings highlight both the potential and the limitations of hybrid approaches and suggest that their effectiveness may depend on the specific context and complexity of the decision environment.