Introduction
Traffic flow forecasting is a fundamental challenge in intelligent transportation systems, requiring an understanding of both spatial dependencies (road network layout) and temporal dynamics (traffic changes over time). Traditional state-of-the-art models often struggle with two primary issues: the bias introduced by manually constructed geographic graphs and the inherent volatility of raw traffic flow data.
FOGS (First-Order Gradient Supervision) addresses these limitations by shifting the prediction target from raw traffic volumes to "trends"—the relative change in flow—and utilizing a custom-learned spatiotemporal graph that captures functional similarities between sensors beyond simple physical proximity.
Methodology
The FOGS framework introduces a two-pronged approach to improve forecasting accuracy: a learning-based graph construction and a novel supervision strategy based on first-order gradients.
1. Learning-based Spatiotemporal Graph
Rather than relying on a static map based on distance, the authors propose learning a representation of sensors that reflects their actual behavior:
- Temporal Feature Vectors: Historical data is averaged into 5-minute slots over a week to create "weekly signatures." Sensors with similar signatures are linked, regardless of their physical distance.
- Custom Random Walks: A sampling strategy integrates the physical road network with temporal similarity to learn sensor embeddings. These embeddings are then used to construct a spatiotemporal correlation graph $G$ based on cosine similarity.
2. First-Order Gradient Supervision (FOGS)
The core innovation of this work is the shift in the learning objective. Instead of predicting the exact number of vehicles, the model predicts the trend ($z$):
$$z_{jt} = \frac{y_{jt} - x_{jT}}{x_{jT}}$$
Where $y$ is the future flow and $x$ is the last observed flow. This transformation is motivated by the statistical properties of the data. Raw traffic flow data often follows an irregular, wide distribution that is difficult for neural networks to fit. In contrast, the distribution of trends is highly concentrated and approximates a Gaussian (Bell Curve) distribution, making it significantly more "learnable."
Experiments and Results
The model was evaluated on four standard benchmarks from the Caltrans Performance Measurement System (PeMS): PEMS03, PEMS04, PEMS07, and PEMS08. These datasets vary in size from 170 to 883 sensors and cover several months of traffic data.
Key Findings
- State-of-the-Art Performance: FOGS significantly outperformed five major baselines, including DCRNN and STGCN. On the PEMS03 dataset, FOGS achieved an improvement in RMSE (Root Mean Square Error) of approximately 14.7% to 19.8%.
- Intuitive Metrics: The model showed particularly strong gains in MAPE (Mean Absolute Percentage Error), which is often more useful for traffic managers as it provides a relative measure of error.
- Backbone Agnostic: FOGS is a training strategy rather than a rigid architecture. By applying the FOGS graph and loss calculation to the existing STFGNN model, the researchers achieved substantial performance boosts without redesigning the underlying neural network.
Conclusion
FOGS demonstrates that how we supervise a model is just as important as the model's architecture. By transforming volatile traffic data into stable trends and learning functional sensor relationships, FOGS provides a more robust framework for spatiotemporal forecasting. Future research will focus on improving performance in sparse networks and handling missing sensor data, which currently limits the effectiveness of gradient-based supervision.
Summary updated