Introduction
Effective traffic management at the operational level, such as adaptive signal control, requires forecasting traffic conditions just seconds into the future. While traditional models rely on data aggregated into 3-to-5-minute intervals, real-time operations demand high-resolution data (0.1–1.0 second intervals). However, this raw data presents significant challenges: it is extremely volatile, non-linear, and heavily skewed toward zeros (binary imbalance), making standard regression models prone to treating actual traffic "jumps" as outliers.
Methodology: Ensembled Kernelized Matrix Completion (EKMC)
To address these challenges, this research introduces the Ensembled Kernelized Matrix Completion (EKMC) framework. This approach shifts the forecasting paradigm from traditional time-series regression to a matrix completion problem, effectively treating future traffic states as missing values in a joint historical-future matrix.
Core Pillars of EKMC
- Matrix Completion: By seeking a low-rank approximation of a joint matrix, the model captures complex correlations between different locations in a network, similar to image inpainting techniques.
- The Kernel Trick: Input data is mapped into a higher-dimensional feature space using kernels, allowing the model to identify non-linear dependencies that are invisible in raw binary occupancy data.
- Adaptive Boosting (AdaBoost): To handle the binary imbalance of high-resolution data, AdaBoost iteratively trains the model. It assigns higher weights to "hard-to-predict" samples, ensuring the model accurately captures infrequent but critical events like the sudden appearance of a vehicle.
- Efficient Optimization: A block coordinate descent algorithm is employed to solve the optimization problem, ensuring the framework can operate in real-time environments.
Experimental Setup
The EKMC model was validated using a real-world high-resolution dataset from downtown Abu Dhabi, UAE. The study area included 11 signalized intersections and two major arterials, providing seven weeks of occupancy data.
Results and Discussion
The proposed EKMC method demonstrated significant improvements over traditional baselines, including Vector Autoregressive (VAR), Support Vector Regression (SVR), and Recurrent Neural Networks (RNN/LSTM).
Performance Metrics
Performance was evaluated using prediction accuracy (1-MAE) and the Skorokhod $M_1$ metric, which specifically measures how well a model captures the timing and shape of traffic "jumps" (transitions from 0 to 1).
Key findings include:
- High Accuracy: EKMC achieved a 1-MAE of over 92% for 1-second horizons and maintained performance above 75% for 2-minute horizons.
- Pattern Recognition: EKMC outperformed VAR and SVR significantly in the $M_1$ metric, indicating a superior ability to recognize the pattern and timing of traffic jumps compared to black-box neural networks.
- Interpretability: Unlike deep learning models, EKMC is an "open-box" system. Engineers can examine learned coefficients to understand which historical features or time points are driving specific predictions.
- Optimal History: Performance peaked at a specific historical data window ($DT=1200$), suggesting that excessive historical data can introduce noise that degrades high-resolution forecasting.
Conclusion and Future Work
EKMC provides a robust, interpretable framework for high-resolution traffic forecasting. While accuracy naturally decays as the prediction horizon increases, the model's ability to handle binary imbalance and non-linear patterns makes it a powerful tool for real-time traffic operations. Future research will focus on optimizing thresholding techniques for binary classification and refining parameter sensitivity across diverse urban environments.
Summary updated