<![CDATA[Short-term traffic condition prediction is essential for supporting responsive monitoring and proactive traffic management within intelligent transportation systems. This study aims to predict multiclass traffic conditions across four categories—normal, heavy, high, and low—for the subsequent 15-minute horizon using an eXtreme Gradient Boosting (XGBoost) model driven by temporal features. The experiment utilized a Kaggle traffic dataset comprising 2,976 chronological observations encompassing timestamp records, categorical vehicle counts, and traffic situation labels. The methodology involved data preprocessing, temporal and cyclical feature engineering, time-ordered data partitioning (2,296 training observations across the first 24 days and 671 testing observations across the final seven days), and hyperparameter optimization. Validation was performed using a three-fold TimeSeriesSplit with an observation gap to prevent temporal data leakage. Experimental results on the holdout test set demonstrated that the proposed XGBoost model achieved an overall accuracy of 75.71%, a weighted F1-score of 73.37%, a macro F1-score of 58.31%, and a balanced accuracy of 56.33%. This performance outperforms the majority baseline (56.18%) and represents an increase of 7.75 percentage points in accuracy over the persistence baseline (67.96%). Class-level analysis revealed strong predictive stability for heavy (F1-score 86.50%) and normal conditions (F1-score 82.21%), although minority classes exhibited lower sensitivity. These findings confirm that integrating XGBoost with engineered temporal features yields an effective approach for 15-minute traffic condition forecasting while rigorously maintaining chronological dependencies.]]>
Copyrights © 2026
