Ensemble machine learning for travel time prediction: An Active learning-assisted approach

• Active learning increases travel time prediction accuracy with less training data. • An ensemble model with active learning outperforms models without active learning. • City-wide case study: 6.72 million trips across the Bangkok road network. • Achieves an MAPE of 10.65 % with the ensemble model trained on 25 % labeled data. • Matches the benchmark model’s accuracy using only 15.62 % of the total data. Accurate travel time prediction is the fundamental challenge in modern transportation systems, affecting real-time navigation, congestion mitigation, and route optimization. However, achieving precise travel time predictions over large transportation networks is challenging due to the dynamic nature of traffic and the high cost of collecting large, labeled datasets. To mitigate these challenges, this study introduces and evaluates an error‑guided active learning strategy that prioritizes high‑error regions for labeling and expands queries to their local spatiotemporal neighborhoods. The framework is evaluated using three models: a Deep Neural Network, an Extreme Gradient Boosting regressor, and an ensemble combination of the two. A case study was conducted using a large-scale dataset of 6.72 million travel time data points from the Bangkok Road network, covering 100 customer locations at 15-minute intervals over a one-week span. Results show the active learning framework substantially reduces labeling without compromising accuracy. A benchmark ensemble model trained on 25 % of randomly selected data attained a Mean Absolute Percentage Error of 10.65 % (12.4 % to 8.9 %). Using the same amount of labeled data, the active learning-enhanced ensemble model achieved a Mean Absolute Percentage Error of 5.95 % (7.2 % to 4.7 %), a roughly 44 % decrease in error. Furthermore, the active learning-enhanced model matched the benchmark’s accuracy using only 15.62 % of the total data, a reduction of almost 40 % in labeling effort while ultimately providing better predictive performance. The proposed method offers a scalable and cost-effective solution for creating high-quality, dynamic travel time predictions that are essential for advancing intelligent transportation and logistics systems.