Hybrid Machine Learning for Crime Prediction in Indonesia toward Society 5.0

Abstract

Crime remains a major social challenge in Indonesia, requiring innovative approaches to enhance prevention and law enforcement. This study proposes a hybrid machine learning framework that integrates the Temporal Fusion Transformer (TFT) for time-series forecasting and Extreme Gradient Boosting (XGBoost) for classification and feature analysis. Using socio-economic and demographic data from the Indonesian Central Bureau of Statistics (2010-2023) across 38 provinces, the framework aims to predict crime incidence and classify crime resolution effectiveness. The results show that TFT effectively captures temporal dependencies, achieving robust forecasting accuracy (R² = 0.9893), while XGBoost delivers high classification performance (Accuracy = 98.87%). Feature importance analysis highlights the dominant role of case resolution rate, government consumption expenditure, school participation rates and life expectancy in shaping crime patterns. Compared to baseline models such as LSTM and Random Forest, the hybrid TFT + XGBoost approach demonstrates superior balance between accuracy, robustness and interpretability. These findings provide actionable insights for policymakers to design data-driven crime prevention strategies, align with Indonesia’s digital transformation agenda, and support the vision of Society 5.0.