Source-linked AI summary
DL-Traff: Survey and Benchmark of Deep Learning Models for Urban Traffic Prediction
Renhe Jiang, Du Yin, Zhaonan Wang, Yizhuo Wang, Jiewen Deng, Hangchen Liu, Zekun Cai, Jinliang Deng, Xuan Song, Ryosuke Shibasaki
TL;DR
Urban traffic prediction lacks organized, comparable evidence across a rapidly growing body of deep-learning models and datasets. DL-Traff surveys the field, defines benchmark tasks and datasets, and evaluates representative models under common settings and metrics. It combines these resources with TensorFlow and PyTorch implementations and is publicly available for reproduction and further use.
Problem
Urban traffic prediction has accumulated many deep-learning models, but evaluations remain confusing because studies use differing datasets, metrics, objectives, and additional data sources.
Method
DL-Traff synthetically surveys deep traffic models and datasets, defines two benchmark tasks, selects open datasets, and evaluates models under the same settings and metrics.
Results
DL-Traff provides a comprehensive benchmark covering grid-based and graph-based models across selected open datasets and prediction tasks.
Takeaways & Limitations
The publicly available DL-Traff resource supports reproducing reported results, using baselines, and launching deep solutions for traffic and other spatiotemporal problems.
Takeaways & Limitations
Multivariate time-series models are only briefly introduced because the paper does not expand their details due to space limitations.
Abstract
from arXiv · showhide
Nowadays, with the rapid development of IoT (Internet of Things) and CPS (Cyber-Physical Systems) technologies, big spatiotemporal data are being generated from mobile phones, car navigation systems, and traffic sensors. By leveraging state-of-the-art deep learning technologies on such data, urban traffic prediction has drawn a lot of attention in AI and Intelligent Transportation System community. The problem can be uniformly modeled with a 3D tensor (T, N, C), where T denotes the total time steps, N denotes the size of the spatial domain (i.e., mesh-grids or graph-nodes), and C denotes the channels of information. According to the specific modeling strategy, the state-of-the-art deep learning models can be divided into three categories: grid-based, graph-based, and multivariate time-series models. In this study, we first synthetically review the deep traffic models as well as the widely used datasets, then build a standard benchmark to comprehensively evaluate their performances with the same settings and metrics. Our study named DL-Traff is implemented with two most popular deep learning frameworks, i.e., TensorFlow and PyTorch, which is already publicly available as two GitHub repositories https://github.com/deepkashiwa20/DL-Traff-Grid and https://github.com/deepkashiwa20/DL-Traff-Graph. With DL-Traff, we hope to deliver a useful resource to researchers who are interested in spatiotemporal data analysis.
1 INTRODUCTION
DL-Traff addresses fragmented evaluation and limited accessibility in deep urban traffic prediction by surveying models, standardizing benchmarks, and releasing reproducible implementations.
- Urban traffic prediction uses large spatiotemporal data from mobile phones, navigation systems, and traffic sensors.
- Traffic data can be represented as a 3D tensor R_T×N×C, with temporal steps, spatial units, and information channels.
- Models are grouped into grid-based, graph-based, and multivariate time-series approaches according to spatial-axis modeling.
- Existing evaluations are difficult to compare because studies use different datasets, metrics, objectives, and additional data sources.
- DL-Traff surveys recent models, selects benchmark tasks and datasets, evaluates state-of-the-art methods under shared settings and metrics, and provides TensorFlow and PyTorch implementations.
2 PROBLEM
DL-Traff defines two benchmark prediction tasks: single-step grid-based inflow/outflow forecasting and multi-step graph-based traffic-speed forecasting.
- The grid-based task predicts inflow and outflow for each mesh-grid in the next time interval.It uses α historical observations and represents each input as X_i ∈ R^(H×W×C), with C=2 for inflow and outflow.
- Together, the tasks cover single-step grid forecasting and multi-step graph forecasting under distinct spatial representations.
- The graph-based task predicts traffic speed across sensor nodes for multiple future steps.It maps α historical observations to β predictions, with X_i ∈ R^(N×C) and C=1 for traffic speed.
3 DATASET
The benchmark selects widely used public urban-traffic datasets spanning taxi and bicycle flows and traffic-sensor measurements from New York, Beijing, Los Angeles, and California.
- The benchmark selects widely used public datasets for urban traffic prediction.The dataset summary enumerates references, sources, and spatial and temporal specifications.
- TaxiBJ contains Beijing taxi inflow and outflow data from four periods between 2013 and 2016.
- BikeNYC contains New York City bicycle inflow and outflow data from April through September 2014.
- METR-LA contains measurements from 207 Los Angeles highway sensors collected over four months in 2012.
- PeMS-BAY and PeMSD7M contain California traffic-sensor data covering 325 and 228 sensors, respectively.PeMS-BAY spans January–May 2017, while PeMSD7M covers May–June 2012 with five-minute weekday sampling.
4 MODEL
DL-Traff organizes urban traffic prediction models around spatial and temporal dependency modeling, covering grid-based, graph-based, and multivariate time-series approaches. It surveys representative architectures and traces how these techniques evolve along spatial and temporal axes.
- 4 MODEL: Urban traffic prediction models address temporal dependencies with LSTM, GRU, 1D CNN, TCN, and attention mechanisms.Attention is also applied to spatial dependency modeling.
- 4 MODEL: Grid-based models use standard convolution over Euclidean mesh-grids, whereas graph-based models use graph convolution with adjacency relations in non-Euclidean road networks.The graph representation may use directed or undirected edges between spatial nodes.
- 4 MODEL: Representative grid-based architectures include ST-ResNet, DMVST-Net, PCRN, STDN, and DeepSTN+, while representative graph-based architectures include STGCN, DCRNN, Graph WaveNet, ASTGCN, and GMAN.The paper plots these architectures in a unified manner for comparison.
- 4 MODEL: Multivariate time-series models evolve along spatial correlation and temporal periodicity but receive less detailed treatment because of space limitations.The paper names LSTNet, TPA-LSTM, GeoMAN, and Transformer as examples.
- 4.1 Roadmap for Grid-Based Model: ST-ResNet models recent observations, daily periodicity, and weekly trends through Closeness, Period, and Trend features.These temporal features were later inherited by models including STDN, DeepSTN+, and ASTGCN.
- 4.2 Roadmap for Graph-Based Model: Graph-based model evolution includes diffusion convolution, multi-graph representations, adaptive or learnable graphs, dynamic structure learning, and attention-augmented GCNs.These developments target richer spatial dependencies beyond a single static graph.
5 EVALUATION
DL-Traff evaluates representative models on grid-based single-step and graph-based multi-step traffic prediction tasks using common experimental settings. The results show strong graph-based performance, while grid-based models improve over baselines without a single dominant model.
- 5 EVALUATION: The benchmark uses HistoricalAverage and CopyLastStep as naive baselines alongside representative deep models.HistoricalAverage averages corresponding historical-day values, while CopyLastStep copies the last one or multiple steps.
- 5.2 Effectiveness Evaluation: Grid-based state-of-the-art models outperform the baselines, but no grid-based model is dominant across the evaluated tasks.STDN performs better generally, while PCRN and DeepSTN+ achieve the lowest MAE on BikeNYC-I and TaxiNYC, respectively.
- 5.2 Effectiveness Evaluation: Grid-based models exhibit practical limitations including slower convergence, overfitting, long training time, large parameter counts, and dependence on multiple data sources.These issues are associated with PCRN, DMVST-Net, STDN, DeepSTN+, and multitask learning models.
- 5.2 Effectiveness Evaluation: DCRNN, Graph WaveNet, and MTGNN rank among the top three across all graph-based datasets, indicating robust performance across datasets.MTGNN and Graph WaveNet obtain most of the highest scores across datasets and metrics.
- 5.2 Effectiveness Evaluation: DCRNN and Graph WaveNet outperform LSTNet in case studies, while all models show time lag during sharp fluctuations, especially at longer horizons.Despite time lag, graph-based models retain better volatility prediction errors.
6 AVAILABILITY AND USABILITY
DL-Traff is publicly available as separate grid-based and graph-based repositories under the MIT License. Its unified scripts and configuration files support architecture inspection, training, testing, and experiment customization.
- 6 AVAILABILITY AND USABILITY: DL-Traff provides separate GitHub repositories for grid-based and graph-based datasets and models under the MIT License.The repositories are DL-Traff-Grid and DL-Traff-Graph.
- 6 AVAILABILITY AND USABILITY: The implementation uses Python with Keras on TensorFlow and PyTorch.Both repositories use popular deep learning frameworks.
- 6 AVAILABILITY AND USABILITY: Users can inspect model architectures, train and test models, customize unified hyperparameters, and access metric and utility functions through the provided scripts and files.MODEL.py, pred_MODEL.py, Param.py, Metrics.py, and Utils.py support these workflows.
7 CONCLUSION
DL-Traff combines a survey of deep learning models and widely used datasets for urban traffic prediction with a standard benchmark on selected open datasets.
- DL-Traff surveys deep learning models and widely used datasets for urban traffic prediction.
- The study builds a standard benchmark to comprehensively evaluate deep traffic models on selected open datasets.
- The survey and benchmark together form the DL-Traff study, which is publicly available through two GitHub repositories.
- DL-Traff is intended as a useful and timely resource for researchers in AI and data science.