Toplu taşımada varış süresi tahmini için karar ağaçları tabanlı makine öğrenmesi modellerinin karşılaştırması

Abstract

Bu çalışma, kentsel toplu taşıma sistemlerinde otobüslerin duraklar arası tahmini varış sürelerini (Estimated Time of Arrival, ETA) daha yüksek doğrulukla öngörebilmek amacıyla karar ağaçları tabanlı makine öğrenmesi algoritmalarının kapsamlı bir karşılaştırmasını sunmaktadır. MBTA tarafından sağlanan “Bus Arrival/Departure Times (2024)” veri seti kullanılarak ham zaman damgaları, sefer kimlikleri ve durak sıraları çok aşamalı bir ön işleme sürecinden geçirilmiş; her seferdeki tüm i < j durak kombinasyonlarını kapsayan durak çifti oluşturulmuştur. Bu yaklaşım, hem kısa mesafeli hem de uzun mesafeli segmentlerin aynı çerçevede modellenmesine olanak tanıyarak seyahat süresi dinamiklerinin daha ayrıntılı temsil edilmesini sağlamıştır. Çalışmada Karar Ağacı, Rastgele Orman, XGBoost, LightGBM ve CatBoost algoritmaları aynı eğitim ve test ayrımı altında değerlendirilmiş; performansları MAE, RMSE ve R² metrikleri ile ölçülmüştür. Üç farklı hat üzerinde yapılan deneylerde CatBoost, tüm modeller içinde en yüksek doğruluk seviyesine ulaşmıştır. Hat 01’de model RMSE = 261.59, MAE = 163.65 ve R² = 0.88; Hat 111’de RMSE = 128.24, MAE = 100.27 ve R² = 0.85; Hat CT2’de ise RMSE = 289.88, MAE = 151.67 ve R² = 0.86 değerlerini elde etmiştir. Eğitim süresi XGBoost ve LightGBM’e kıyasla daha uzun olmakla birlikte, CatBoost’un doğruluk, istikrar ve kategorik değişken işleme kapasitesiyle ETA tahmininde uygulanabilir bir üst seviye performans sunduğu görülmüştür.

Keywords

• Varış Süresi Tahmini
• Karar Ağaçları
• Rastgele Orman
• LightGBM
• CatBoost
• XGBoost

Comparison of decision tree-based machine learning models for arrival time prediction in public transportation

Abstract

This study presents a comprehensive comparison of decision tree-based machine learning algorithms to predict bus arrival times (ETAs) between stops with higher accuracy in urban public transportation systems. Using the "Bus Arrival/Departure Times (2024)" dataset provided by the MBTA, raw timestamps, trip IDs, and stop sequences were subjected to a multi-stage preprocessing process. A stop-pair structure encompassing all stop combinations i < j on each trip was created. This approach allows for the modelling of both short-distance and long-distance segments within the same framework, providing a more detailed representation of travel time dynamics. In this study, the Decision Tree, Random Forest, XGBoost, LightGBM, and CatBoost algorithms were evaluated under the same training-testing scenario, and their performance was measured using MAE, RMSE, and R² metrics. In experiments conducted on three different lines, CatBoost achieved the highest accuracy among all models. In Line 01, the model obtained RMSE = 261.59, MAE = 163.65, and R² = 0.88; in Line 111, RMSE = 128.24, MAE = 100.27, and R² = 0.85; and in Line CT2, RMSE = 289.88, MAE = 151.67, and R² = 0.86. Although the training time is longer compared to XGBoost and LightGBM, CatBoost has been shown to offer a feasible high-level performance in ETA estimation with its accuracy, stability, and categorical variable processing capacity.

Keywords

• Estimated Time of Arrival
• Decision Trees
• Random Forest
• LightGBM
• CatBoost
• XGBoost

References

1. Büchel, B. ve Corman, F. (2020). Review on statistical modeling of travel time variability for road-based public transport. Frontiers in Built Environment, 6, Article 70.
2. Vlahogianni, E. I., Karlaftis, M. G. ve Golias, J. C. (2014). Shortterm traffic forecasting: Where we are and where we’re going. Transportation Research Part C, 43, 3–19.
3. Jenelius, E. ve Koutsopoulos, H. N. (2013). Travel time estimation for urban road networks using low-frequency probe vehicle data. Transportation Research Part B: Methodological, 53, 64–81.
4. Sun, D., Luo, H., Fu, L., Liu, W., Liao, X. ve Zhao, M. (2007). Predicting bus arrival time on the basis of global positioning system data. Transportation Research Record, 2034(1), 62–72.
5. Fan, W. ve Gurmu, Z. (2015). Dynamic travel time prediction models for buses using GPS data.Transportation Research Part C, 59, 18–32.
6. Moosavi, S. M. H., Aghaabbasi, M., Yuen, C. W. ve Jahed Armaghani, D. (2023). Evaluation of applicability and accuracy of bus travel time prediction in high and low-frequency bus routes using tree-based machine learning techniques. Journal of Soft Computing in Civil Engineering, 7(2), 74–97.
7. Hastie, T., Tibshirani, R. ve Friedman, J. (2009). The Elements of Statistical Learning: Data Mining, Inference, and Prediction (2nd ed.). Springer.
8. Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32.

9. Chen, T. ve Guestrin, C. (2016). XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 785–794). ACM.
10. Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q. ve Liu, T.-Y. (2017). LightGBM: A highly efficient gradient boosting decision tree. In Advances in Neural Information Processing Systems (NeurIPS 2017) (pp. 3149–3157).
11. Prokhorenkova, L., Gusev, G., Vorobev, A., Dorogush, A. V. ve Gulin, A. (2018). CatBoost: unbiased boosting with categorical features. In Advances in Neural Information Processing Systems (NeurIPS 2018) (pp. 6638–6648).
12. Cramér, H. (1946). Mathematical Methods of Statistics. Princeton University Press.
13. Cats, O. (2011). Dynamic Modelling of Transit Operations and Passenger Decisions. TU Delft.
14. Okutucu, H. ve Kaya, Ş. M. (2025). Traffic Density Estimation with IoT Edge Computing In Smart City: A Case Study. Anadolu Bil Meslek Yüksekokulu Dergisi, 20(72), 189-215.
15. Esmaili Jobani, A. ve Kaya, Ş. M. (2025). Hybrid IoT and AIbased Solution for Energy Management in Data Centres under Various Climate Conditions. Anadolu Bil Meslek Yüksekokulu Dergisi, 20(72), 107-124.
16. Kaya, Ş. M., Güneş, A. ve Erdem A. (2021). "A Smart Data Pre- Processing Approach by Using ML Algorithms on IoT Edges: A Case Study," 2021 International Conference on Artificial Intelligence of Things (ICAIoT), Nicosia, Turkey, 2021, pp. 36-42.
17. İşler, B., Kaya, Ş.M. ve Kılıç, F.R. (2025). Fog-Enabled Machine Learning Approaches for Weather Prediction in IoT Systems: A Case Study. Sensors 2025, 25, 4070.