TL;DR
This paper provides a comprehensive roadmap for Short-Term Traffic Prediction (STTP) using deep learning, categorizing how researchers represent complex spatiotemporal data and tracking the evolution of neural architectures from early unsupervised models to modern graph-based and meta-learning systems. It serves as both a technical primer and a directory of the most influential datasets and methodologies in the field.
What Problem Does This Paper Address?
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 "spatiotemporal" dependencies. This means what happens at one location is influenced by its neighbors (spatial) and by what happened minutes or hours ago (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. Researchers use different ways to format their data and different architectures to process it, making it nearly impossible to compare results across studies. This paper addresses this fragmentation by providing a unified taxonomy for data representation and model selection.
Methods
The authors categorize the STTP landscape into four primary perspectives:
- Input Data Representation: This is perhaps the paper's most critical contribution. They classify spatial data into:
- Stacked Vectors: Simple lists of data points.
- Grid Representations: Treating a city map like an image (2D pixels), ideal for Convolutional Neural Networks (CNNs).
- Graphs: Treating road segments as nodes and intersections as edges, which captures the actual topology of a road network better than a grid.
- Temporal Dependencies: They distinguish between simple sequentiality (what happened just before) and periodicity (daily or weekly cycles).
- DNN Taxonomy: The paper tracks five "generations" of models:
- 1st Gen: Unsupervised models like Restricted Boltzmann Machines (RBMs).
- 2nd Gen: Basic supervised models (MLP, CNN, RNN/LSTM).
- 3rd Gen: Hybrids (e.g., using a CNN for spatial features and an RNN for temporal ones).
- 4th Gen: Graph Convolutional Networks (GCNs), which handle non-Euclidean road networks.
- 5th Gen: Advanced techniques like Attention mechanisms, Reinforcement Learning, and Meta-learning.
Datasets
The survey compiles a list of 16 popular public datasets. Key benchmarks include:
- PeMS: Data from nearly 40,000 individual detectors on California freeways. It is the most widely used dataset for traffic flow and speed.
- TaxiBJ & TaxiNYC: Trajectory and trip record data from Beijing and New York City taxis, primarily used for demand prediction.
- BikeNYC & BikeCHI: Bike-sharing data used to predict crowd flow and station-level demand.
- Didi: Ride-hailing data from Chengdu, China.
The authors note that while PeMS is a "gold standard" for freeways, there is still a desperate need for a diverse, "ImageNet-style" benchmark for urban traffic.
Key Insights and Ideas
- Inductive Bias is Key: The choice of architecture imposes a "prior" on the problem. For example, using a CNN implies that "near things are more related than distant things" in a grid-like fashion. If your road network doesn't fit a grid, this bias can actually hurt performance.
- Entity Type Dictates Representation: The authors find that "human" traffic (pedestrians, bikes) is often successfully modeled using grid/image-like representations because humans move somewhat freely. However, "vehicle" traffic is strictly constrained to roads, making graph-based representations far more efficient and accurate.
- The "No-Free-Lunch" Theorem: No single DNN architecture wins every time. The best model depends heavily on the specific spatiotemporal characteristics of the target city and the quality of the available data.
Key Results
- Dominance of PeMS: Out of the surveyed papers, 27 used the PeMS dataset, highlighting its role as the closest thing the field has to a standard benchmark.
- Evolution Toward Graphs: The survey shows a clear trend moving from 2nd-generation RNNs/LSTMs toward 4th-generation GCNs. This shift reflects the community's realization that road networks are fundamentally graphs, not grids.
- Hybrid Success: Third-generation hybrid models (CNN + RNN) consistently outperform single-architecture models by capturing both spatial "images" of traffic and temporal "sequences" of flow.
Limitations and Future Work
- Code Availability: A major hurdle identified is the lack of open-source code. Many high-performing models are described in papers but cannot be easily replicated or built upon.
- Standardization: The lack of a standard benchmark for urban (non-freeway) traffic remains a significant gap.
- Future Mobilities: The authors suggest that STTP must evolve to handle "Future Mobilities," such as drones (which require 3D spatial modeling) and autonomous vehicles (which create a mixed environment of human and machine drivers).
- Meta-Learning: The paper highlights meta-learning as a promising frontier for "transferring" traffic knowledge from a data-rich city (like NYC) to a data-scarce city.
Summary updated