Data labeling is a critical—and often costly—step in building supervised machine learning models, especially in domains like road traffic safety, where only a small subset of observations can be manually annotated, and class imbalance is severe. We propose a hybrid, semi-supervised labeling pipeline that combines three strategies: (1) a rule-based voting ensemble, in which domain experts define attribute-specific threshold rules (weak classifiers) whose outputs are aggregated by majority vote over a small manually labeled seed set, (2) fuzzy c-means clustering to assign soft labels in a complementary, unsupervised manner, and (3) fuzzy sets, in which a fuzzy inference system based on threshold rules and penalty values is built to determine the risk level of a single observation. To remedy label sparsity, we augment rare classes with synthetic examples following expert-driven risky patterns and balance the final annotations via SMOTE. On a real-world driving dataset (23,152 synthetically enriched observations; 21,172 unlabeled), our voting ensemble achieves 82% labeling accuracy on held-out expert labels-preserving the expected “descending” class distribution from low to very high risk—while fuzzy clustering often misclassifies high-risk cases. Our approach yields a fully labeled, balanced dataset of 50,612 instances ready for downstream training with minimal manual effort and clear interpretability.