Abstract
In modern transportation systems, an enormous amount of traffic data is generated every day. This has led to rapid progress in short-term traffic prediction (STTP), in which deep learning methods have recently been applied. In traffic networks with complex spatiotemporal relationships, deep neural networks (DNNs) often perform well because they are capable of automatically extracting the most important features and patterns. In this study, we survey recent STTP studies applying deep networks from four perspectives: input data representation, DNN techniques, application areas/datasets, and standard benchmarks. Finally, we suggest challenging issues and possible future research directions in STTP.
The Challenge of Spatiotemporal Traffic Data
Predicting traffic—whether it's vehicle flow on a highway, taxi demand in a city, or pedestrian crowds in a subway—is notoriously difficult because it involves complex "spatiotemporal" dependencies. This means what happens at one location is influenced by its neighbors (spatial) and by historical patterns (temporal). While classical statistical models like ARIMA have been used for decades, they struggle with the non-linear, "messy" nature of real-world traffic. Deep Neural Networks (DNNs) are better suited for this, but the field has been fragmented by varying data formats and architectures.
Taxonomy of Input Data Representation
A critical contribution of this survey is the classification of how researchers represent spatial data. The choice of representation often dictates the choice of neural architecture:
- Stacked Vectors: Simple lists of data points. These are easy to implement for simple, linear road segments but fail as network complexity increases.
- Grid Representations: Treating a city map like an image (2D pixels), which is ideal for Convolutional Neural Networks (CNNs).
- Graphs: Treating road segments as nodes and intersections as edges. This captures the actual topology of a road network far better than a grid, especially for vehicle traffic constrained to specific paths.
The Evolution of Neural Architectures
The paper tracks five "generations" of models, showing a clear progression toward capturing more complex relationships:
- Unsupervised Models: Early networks like Restricted Boltzmann Machines (RBMs).
- Basic Supervised Models: Standard MLP, CNN, and RNN/LSTM architectures.
- Hybrid Models: Combining architectures (e.g., CNNs for spatial features and RNNs for temporal sequences).
- Graph Convolutional Networks (GCNs): Specifically designed to handle non-Euclidean road networks.
- Advanced Techniques: The current frontier, including Attention mechanisms, Reinforcement Learning, and Meta-learning.
Key Insights: Inductive Bias and Spatial Context
The survey highlights that the choice of architecture imposes an "inductive bias" on the problem. For example, using a CNN implies that "near things are more related than distant things" in a grid-like fashion. However, traffic is often influenced by the functional nature of the area, such as the flow between residential zones and business districts.
Furthermore, the authors find that "human" traffic (pedestrians, bikes) is often successfully modeled using grid representations because movement is relatively free. In contrast, "vehicle" traffic is strictly constrained to roads, making graph-based representations far more efficient and accurate.
Benchmarking and Future Directions
The survey compiles 16 popular public datasets, noting that the PeMS dataset (California freeways) has become the closest thing to a standard benchmark. However, a significant gap remains: the lack of a diverse, "ImageNet-style" benchmark for complex urban traffic.
Looking forward, the authors suggest that STTP must evolve to handle "Future Mobilities," such as drones (requiring 3D spatial modeling) and autonomous vehicles. Additionally, Meta-learning is identified as a promising frontier for transferring traffic knowledge from data-rich cities to data-scarce regions, potentially solving the cold-start problem in urban planning.
Summary updated