Deep Learning on Traffic Prediction: Methods, Analysis, and Future Directions

Yin, Xueyan; Wu, Genze; Wei, Jinze; Shen, Yanming; Qi, Heng; Yin, Baocai · 2022 · OpenAlex-citations

DOI: 10.1109/tits.2021.3054840

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper presents a comprehensive survey of deep learning-based approaches for traffic prediction, addressing the critical challenge of modeling complex, dynamic spatio-temporal dependencies in road networks. Traffic prediction is essential for intelligent transportation systems, enabling applications such as congestion mitigation, route planning, and vehicle dispatching. The authors identify two primary challenges: complex spatial dependencies, where the influence of different locations on a target position varies dynamically, and dynamic temporal dependencies, characterized by non-linear changes and periodicity in traffic data. The study aims to provide a systematic taxonomy of existing methods, organize public datasets, evaluate model performance, and outline future research directions. The authors categorize traffic prediction methods into classical approaches and deep learning-based models. Classical methods include statistical techniques like ARIMA and VAR, which are limited by assumptions of linearity and small dataset sizes, and traditional machine learning methods such as Support Vector Regression and Random Forests, which handle non-linear relationships but lack robust spatio-temporal modeling. The survey focuses extensively on deep learning architectures, detailing how Convolutional Neural Networks (CNNs) extract spatial features from grid-structured data, while Graph Convolutional Networks (GCNs) model non-Euclidean graph structures inherent in road networks. The paper distinguishes between spectral-based GCNs, which use graph signal processing, and spatial-based GCNs, which aggregate neighbor features. Additionally, it reviews Recurrent Neural Networks (RNNs), including LSTM and GRU variants, for temporal dependency modeling, and discusses the integration of attention mechanisms to dynamically weigh spatial and temporal correlations. The survey provides a detailed analysis of key techniques, including diffusion convolution for capturing bidirectional traffic flow and attention mechanisms for adapting to varying impacts of neighboring regions over time. The authors organize widely used public datasets and external factors, such as weather and events, that influence traffic patterns. They conduct extensive comparative experiments on real-world public datasets to evaluate the performance of different deep learning models, identifying effective components for spatio-temporal modeling. The study covers various application tasks, including traffic flow, speed, demand, travel time, and occupancy prediction. The significance of this work lies in its role as the first comprehensive survey covering deep learning in traffic prediction from multiple perspectives, including methods, applications, datasets, and experimental analysis. By providing a structured taxonomy and evaluating state-of-the-art approaches, the paper offers a valuable resource for researchers to understand the evolution from classical to deep learning methods. It highlights the limitations of current solutions, such as computational complexity and the challenge of selecting relevant historical observations, and proposes promising future directions. This systematic review facilitates further advancements in intelligent transportation systems by clarifying the strengths and weaknesses of existing deep learning architectures in handling complex spatio-temporal data.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• anticipation