Comparison of Oversampling Techniques in Prediction Judicial Decisions of Divorce Trials in Family Courts

This study delves into predicting decisions in family courts, explicitly addressing the challenges of imbalanced datasets in divorce trials. We evaluated four oversampling techniques to enhance machine learning model accuracy in these contexts-SMOTE, Random Over, GAN, and ADASYN. Historical Surabaya’s Religious Court case records were preprocessed and divided into training and testing sets. Alongside TF-IDF vectorization, we incorporated Indo BERT to prepare the textual data for machine learning analysis. These oversampling methods were employed before model training to counteract class imbalance. An extensive grid search identified optimal hyperparameters, refining the performance of models like Support Vector Machines, Random Forest, and Naive Bayes. Models were assessed through accuracy, precision, recall, and F1-score metrics to determine the most effective oversampling and machine learning model combination. The most notable results were achieved with Random Forest paired with SMOTE, achieving a precision of 0.82%, a recall of 0.81%, an F1-Score of 0.80%, and an accuracy of 0.81%. This study highlights the significance of tailored oversampling strategies and advanced machine learning for accurate, data-driven decisions in family courts, emphasizing thorough data preparation and meticulous model tuning.