Machine Learning-Based Prediction of Divorce Verdicts Using Posita Data and Imbalanced Data Handling: A Case Study in Padang Sidempuan

Abstract

This study aims to develop a predictive model for divorce verdicts ("Granted" or "Rejected") in the Religious Courts of Indonesia using machine learning techniques. The dataset consists of 2,026 finalized divorce cases from the Religious Court of Padang Sidempuan between 2018 and 2025, incorporating structured variables and posita—narrative texts describing the plaintiff’s reasons for divorce. Keyword-based feature extraction was applied to transform these texts into interpretable indicators. To handle class imbalance, Synthetic Minority Over-sampling Technique (SMOTE) was implemented on the training data. Six classical machine learning algorithms were evaluated: Decision Tree, Naïve Bayes, K-Nearest Neighbors, Random Forest, LightGBM, and XGBoost. Performance was measured using accuracy, precision, recall, F1-score, F2-score, and AUC. The results indicate that Naïve Bayes achieved the highest recall (100%) for the “Granted” class, while LightGBM and XGBoost demonstrated the most balanced performance across both classes. Feature importance analysis revealed that mediation outcomes, domestic violence, and economic hardship were among the most influential factors in determining verdicts. The study highlights the applicability of interpretable machine learning in legal decision support and discusses limitations such as the single-court scope and challenges in predicting minority class outcomes. Future work may explore multi-jurisdictional data, deep learning approaches, and domain-specific embeddings for enhanced performance.