Smart Organization of Imbalanced Traffic Datasets for Long-Term Traffic Forecasting

Area of Science:

• Traffic engineering
• Machine learning
• Data science

Background:

• Traffic speed prediction is crucial for urban efficiency and pollution reduction.
• Existing research often overlooks imbalanced datasets, where low traffic speeds are underrepresented.
• Traffic data is influenced by temporal factors like time of day and weather (month).

Purpose of the Study:

• To address the challenge of imbalanced datasets in traffic speed prediction.
• To develop novel data organization techniques for improved minority class representation.
• To evaluate the effectiveness of these techniques across various machine learning models.

Main Methods:

• Devised Hour-wise and Month-wise Pattern Organization techniques using temporal factors.
• Proposed a Speed-wise Pattern Organization strategy considering traffic volatility.
• Evaluated strategies using Long Short-Term Memory (LSTM), Gated Recurrent Unit networks (GRUs), Bi-directional LSTM, and Convolutional Neural Networks (CNNs).

Main Results:

• GRU models achieved the best performance with a Mean Absolute Percentage Error (MAPE) of 13.51%.
• The proposed methodologies improved overall model accuracy by approximately 4%.
• Prediction accuracy for low-traffic speed scenarios was augmented by 11.2%.

Conclusions:

• The developed data organization techniques effectively improve traffic speed prediction, particularly for imbalanced datasets.
• These pre-processing methodologies enhance the performance of various machine learning models.
• The study offers a robust approach to tackle data imbalance in traffic prediction for better urban mobility.