Model Prediksi Keberhasilan Penuntutan Perkara Korupsi Menggunakan Explainable Machine learning

Abstract

The integration of artificial intelligence in legal analytics often encounters substantial friction due to the inherent "black-box" nature of advanced predictive models, which conflicts with the strict requirements for transparency and accountability in judicial decision-making. To address this limitation, this study develops a robust transdisciplinary framework using Explainable Machine learning (XML) to predict prosecution outcomes in corruption cases while systematically decomposing the underlying causal mechanisms. Utilizing an extensive dataset of inkracht corruption cases, six machine learning algorithms—ranging from linear baselines to advanced boosting architectures—were trained, optimized via Bayesian optimization, and evaluated using rigorous performance and calibration metrics. The empirical results demonstrate that the XGBoost model outperforms competing architectures, achieving an Accuracy of 0.92, an F1-Score of 0.91, a Matthews Correlation Coefficient (MCC) of 0.84, and an optimal probability calibration with a Brier Score of 0.06. McNemar’s test confirms that this predictive superiority is statistically significant ($p < 0.05$). Beyond raw predictive performance, global and local interpretability via post-hoc methods (SHAP and LIME) reveals that the presence of electronic evidence and the number of admissible pieces of evidence serve as the most critical positive determinants, whereas high case complexity exerts a strong negative influence on prosecution success, directly validating judicial burden-of-proof theories. This dual-granular explanation bridges the gap between computational law and prosecutorial discretion. Ultimately, this framework provides a scientifically validated, non-prescriptive decision support system capable of mitigating prosecution risks, reducing systemic disparities, and supporting data-driven governance within judicial administrations.