CLASSIFICATION OF PRE-MARITAL SEX AMONG ADOLESCENTS USING COST-SENSITIVE WEIGHTED RANDOM FOREST ON IMBALANCED DATA

Abstract

Adolescents are a critical developmental phase characterized by significant physical, cognitive, and psychosocial changes, which increase vulnerability to risky behaviors, including premarital sexual activity. Such behavior may lead to serious consequences, such as unintended pregnancy, unsafe abortion, sexually transmitted infections, and HIV/AIDS. This study aims to classify premarital sexual behavior among adolescents using demographic, knowledge, and behavioral factors. The data were obtained from the 2019 Program Performance and Accountability Survey (SKAP), consisting of approximately 5,300 weighted adolescent respondents in East Java. The dataset exhibits extreme class imbalance, with only 0.3% of adolescents reporting premarital sexual behavior. To address this issue, three classification methods were applied: standard Random Forest (RF), Weighted Logistic Regression (WLR), and Cost-Sensitive Weighted Random Forest (CSWRF). Model performance was evaluated using 5-fold cross-validation and multiple metrics, including accuracy, sensitivity, specificity, precision, F1-score, G-Means, and Area Under the Curve (AUC). The results indicate that although the standard RF achieved very high accuracy (above 99%), it failed to identify the minority class, resulting in undefined specificity and poor discriminative ability (AUC ≈ 0.83). The WLR model improved classification balance, achieving reasonable specificity (0.8667) and AUC (0.9459), but produced more false positives. In contrast, the CSWRF model demonstrated the best overall performance, with high AUC values (above 0.95), improved specificity, and a better balance between sensitivity and specificity, indicating its effectiveness in handling highly imbalanced data. Variable importance analysis using Mean Decrease Accuracy (MDA) identified Knowledge of Adolescent Reproductive Health (ARH), residence type, and behavioral factors such as dating experience as the most influential predictors. Despite these findings, this study is limited by potential underreporting bias and exclusion of socio-cultural and psychological variables. Future research should incorporate additional predictors and explore advanced modeling approaches to improve classification performance.