A New Evaluation Metrics and Hybrid Deep Learning Model to Enhance Predictive Performance of Traffic Flow

Mouna MNIF, Sadok Bouamama

Keywords: Deep Learning, Gated Recurrent Unit, Long Short-Term Memory, Hybrid model, Traffic Flow Prediction, Evaluation metrics, Balancing Coefficients, Gain metric.

Abstract: Traffic flow prediction is essential in transportation management, as it can help reduce congestion and optimize the planning of transport. With the emergence of deep learning techniques, several models have been developed for traffic flow prediction. In this research paper, we proposed new evaluation metrics to calculate the gain of performance prediction models, based on the Root Mean Square Error (RMSE), Mean Squared Error (MSE), and R-squared (R²) scores with balancing coefficients. Then, we proposed a new hybrid Long Short-Term Memory-Gated Recurrent Unit (LSTM-GRU) model. The experimental results of this study emphasize the effectiveness of deep learning models in predicting traffic flow. The hybrid models combining GRU (Gated Recurrent Unit) and LSTM (Long Short-Term Memory) show promising performances in forecasting tasks. The hybrid GRU-LSTM (Gated Recurrent Unit-Long Short-Term Memory) emerges as the best compromise vs. LSTM, GRU, and LSTM with the Wadam optimizer. This model exhibits stable performance across the train, validation, and test sets, indicating it does not suffer from overfitting, as proved by the gain evaluation metrics. Besides, it provides a good balance of accuracy, stability, and computational efficiency.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 13986