A Comparative Evaluation of SMOTE Placement Strategies for Multi-Class Traffic Complaint Classification Using SVM
Faradilla Nur Azizah, Agus Sulistya, Berlian Rahmy Lidiawaty · 2025
The rapid growth of traffic complaints on social media presents challenges for authorities to manage and prioritize responses effectively. This study proposes a multiclass classification approach to categorize complaint texts into three urgency levels: low, medium, and fatal. A dataset of 450 tweets collected from the @e100ss transportation account on Platform$\mathbf{X}$(formerly Twitter) was used. The classification was performed using the Support Vector Machine (SVM) algorithm with multiple kernel functions. To address the class imbalance, two SMOTE (Synthetic Minority Over-sampling Technique) placement strategies—before and after train-test splitting—were evaluated. The best performance was achieved using the RBF kernel combined with SMOTE-before-splitting, yielding 85.43% accuracy and high F1-scores across all classes. In contrast, applying SMOTE after the split significantly reduced the recall for fatal complaints. The model's consistency was confirmed through 5-fold crossvalidation with an average F1-macro score of 86%. These findings demonstrate the importance of SMOTE placement in urgencyaware classification systems for imbalanced social media data and offer practical implications for improving traffic complaint prioritization.