Are state-of-the-art deep learning traffic prediction models truly effective?
image: All of the baseline models achieve excellent performance in predicting high speed while performing extremely poorly in predicting lower ones. Specifically, even if the prediction horizon is 60 mins, a quite further future, the RMSE is as low as 4 and 2 for METR-LA and PEMS-BAY when the average speed is greater than 60 mph, respectively. However, RMSE is around 20 for predicting speeds under 20 mph, which suggests that the model almost randomly generates the prediction results. Correspondingly, MAPE is nearly 100% in this case.
Credit: Communications in Transportation Research
Accurate and efficient traffic speed prediction is crucial for improving road safety and efficiency. With the emerging deep learning and extensive traffic data, data-driven methods are widely adopted to achieve this task with increasingly complicated structures and progressively deeper layers of neural networks. Despite the design of the models, they aim to optimize the overall average performance without discriminating against different traffic states. However, the fact is that predicting the traffic speed under congestion is normally more important than the one under free flow since the downstream tasks, such as traffic control and optimization, are more interested in congestion rather than free flow. Unfortunately, most of the state-of-the-art (SOTA) models do not differentiate the traffic states during training and evaluation. To this end, we first comprehensively study the performance of the SOTA models under different speed regimes to illustrate the low accuracy of low-speed prediction. We further propose and design a novel Congestion-Aware Sparse Attention transformer (CASAformer) to enhance the prediction performance under low-speed traffic conditions. Specifically, the CASA layer emphasizes the congestion data and reduces the impact of free-flow data. Moreover, we adopt a new congestion adaptive loss function for training to make the model learn more from the congestion data. Extensive experiments on real-world datasets show that our CASAformer outperforms the SOTA models for predicting speed under 40 mph in all prediction horizons.
They published their study on 10 April 2025, in Communications in Transportation Research.
Effective traffic speed prediction is a cornerstone of intelligent transportation systems, where the main tasks include but are not limited to traffic control, optimized routing, and congestion management. Existing deep learning models, while excelling in overall accuracy, perform poorly in low-speed predictions. Thus, such limitation may degrade their benefit for congestion-focused applications. This study highlights the limitations of SOTA models like Spatiotemporal Graph Neural Networks (STGNNs) and Transformers under various traffic regimes, revealing a systematic bias toward free-flow conditions.
CASAformer is specifically designed to improve the accuracy of low-speed traffic prediction. It introduces three main modules.
1. Embedding Layers
Embedding layers are responsible for encoding the raw data to include multiple different features. Specifically, the embedding layers will output the following types of embedding:
- Historical Data Embedding: Encodes traffic speed data collected from sensors over historical time frames, projecting it into a latent feature space.
- Periodicity Embedding: Captures time-dependent patterns by integrating time-of-day and day-of-week information.
- Adaptive Embedding: Learns spatial-temporal dependencies unique to each sensor, enhancing the model’s ability to handle diverse traffic conditions.
The embeddings are concatenated to create an input representation that provides a rich and comprehensive feature set for further processing.
2. Temporal Self-Attention
Temporal attention mechanisms are applied along the time axis for each traffic sensor. The model computes attention scores to determine the relative importance of historical time steps, effectively capturing temporal dependencies in traffic dynamics.
3. Spatial Congestion-Aware Sparse Attention (CASA) Layers
CASA layers are the core of our model. Rather than applying global attention and computing attention scores for every pair of sensors, we propose to “pay more attention” to the congested sensors. Specifically, sparse attention, including congestion attention and informative attention mechanisms, are carefully designed and implemented to overcome the biased information introduced by a majority of free-flow data.
- Congestion Attention: Identifies sensors under congestion (defined by a speed threshold) and assigns them higher attention weights.
- Informative Attention: Selects a subset of non-congested sensors based on diversity in historical speed data to ensure that the model captures relevant features while reducing noise.
By focusing only on congested and informative nodes, CASA reduces the computational complexity compared to traditional global attention mechanisms. The outputs of the CASA layers are combined to form a sparse attention matrix, which can both emphasize congested regions and efficiently summarize spatial data.
A novel loss function is also proposed to assign higher weights to congested traffic states and make the model learn more from low-speed condition data. The function dynamically adapts based on speed thresholds to mitigate the data imbalance issue.
CASAformer bridges a critical gap in low-speed prediction, offering enhanced tools for congestion-focused traffic management. Its sparse architecture provides a scalable solution for ITS deployments. However, the following limitations still remain. Performance variability across datasets highlights the need for robust parameter generalization. Trade-offs in high-speed predictions call for further optimization to maintain balanced performance. Future work will explore cross-dataset adaptability and hybrid architectures to integrate CASA layers with other SOTA models.
About Communications in Transportation Research
Communications in Transportation Research was launched in 2021, with academic support provided by Tsinghua University and China Intelligent Transportation Systems Association. The Editors-in-Chief are Professor Xiaobo Qu, a member of the Academia Europaea from Tsinghua University and Professor Shuai’an Wang from Hong Kong Polytechnic University. The journal mainly publishes high-quality, original research and review articles that are of significant importance to emerging transportation systems,aiming to serve as an international platform for showcasing and exchanging innovative achievements in transportation and related fields, fostering academic exchange and development between China and the global community.
I It has been indexed in SCIE, SSCI, Ei Compendex, Scopus, DOAJ, TRID and other databases.
It was selected as Q1 Top Journal in the Engineering and Technology category of the Chinese Academy of Sciences (CAS) Journal Ranking List. In 2022, it was selected as a High-Starting-Point new journal project of the “China Science and Technology Journal Excellence Action Plan”. In 2024, it was selected as the Support the Development Project of “High-Level International Scientific and Technological Journals”. The same year, it was also chosen as an English Journal Tier Project of the “China Science and Technology Journal Excellence Action Plan PhaseⅡ”. The 2024 IF is 14.5, ranking in the Top1 (1/61, Q1) among all journals in "TRANSPORTATION" category. Tsinghua University Press will cover the open access fee for all published papers in 2025.