Graph autoencoder with mirror temporal convolutional networks for traffic anomaly detection

Zhiyu Ren, Xiaojie Li, Jing Peng, Ken Chen, Qushan Tan, Xi Wu, Canghong ShiScientific 2024 · DOI

Abstract

Traffic time series anomaly detection has been intensively studied for years because of its potential applications in intelligent transportation. However, classical traffic anomaly detection methods often overlook the evolving dynamic associations between road network nodes, which leads to challenges in capturing the long-term temporal correlations, spatial characteristics, and abnormal node behaviors in datasets with high periodicity and trends, such as morning peak travel periods. In this paper, we propose a mirror temporal graph autoencoder (MTGAE) framework to explore anomalies and capture unseen nodes and the spatiotemporal correlation between nodes in the traffic network. Specifically, we propose the mirror temporal convolutional module to enhance feature extraction capabilities and capture hidden node-to-node features in the traffic network. Morever, we propose the graph convolutional gate recurrent unit cell (GCGRU CELL) module. This module uses Gaussian kernel functions to map data into a high-dimensional space, and enables the identification of anomalous information and potential anomalies within the complex interdependencies of the traffic network, based on prior knowledge and input data. We compared our work with several other advanced deep-learning anomaly detection models. Experimental results on the NYC dataset illustrate that our model works best compared to other models for traffic anomaly detection.

Introduction

Urban transportation networks generate massive streams of spatiotemporal data. Detecting "anomalies"—events like accidents or sudden congestion that deviate from normal patterns—is critical for safety and efficiency. However, traffic networks are complex, moving graphs where sensors (nodes) are connected by roads (edges). Detecting anomalies is difficult due to high dimensionality, data sparsity, and dynamic dependencies where the relationship between nodes changes over time (e.g., rush hour bottlenecks).

Existing tools like LSTMs or standard Temporal Convolutional Networks (TCNs) often fail to capture long-term correlations and hidden node-to-node features. To address this, the Mirror Temporal Graph Autoencoder (MTGAE) framework was developed to better model these evolving dynamic associations.

Methodology: The MTGAE Framework

The MTGAE utilizes an unsupervised encoder-decoder architecture. The model is trained on "normal" traffic data to learn how to reconstruct it. When an anomaly occurs, the reconstruction error increases, signaling a detection.

The overall architecture of the MTGAE framework, featuring an encoder-decoder structure.

The framework's effectiveness stems from two primary innovations: the Mirror Temporal Convolutional Module (MTCM) and the Graph Convolutional Gated Recurrent Unit (GCGRU) cell.

Core Innovations: MTCM and GCGRU

Standard TCNs use causal convolutions that only look backward in time, limiting their receptive field. The MTCM performs a "mirror flip" on the input sequence, effectively doubling the model's perceptual field and allowing it to capture hidden relationships without losing historical context.

The GCGRU cell integrates Graph Convolutional Networks (GCNs) directly into the gating units of a standard GRU. Furthermore, it employs Gaussian kernel functions to map traffic data into a high-dimensional space. This mapping acts as a filter, making it easier to separate anomalous signals from normal patterns that might appear similar in lower dimensions.

Detailed view of the Mirror Temporal Convolutional Module (MTCM) and the Graph Convolutional Gated Recurrent Unit (GCGRU) cell.

Experimental Evaluation

The model was evaluated across three diverse datasets:

  1. NYC Taxi Dataset: Six months of trip data with synthetic spatial and temporal anomalies injected to test detection precision.
  2. PEMS-BAY: Real-time freeway sensor data from the California Bay Area, used for ablation studies.
  3. DGraphFin: A massive financial dataset (3.7 million nodes) used to test the model’s cross-domain generalization.

Key Results

  • State-of-the-Art Accuracy: On the NYC dataset, MTGAE achieved an AUC of 1.0 across several experimental setups, outperforming baselines like GraphSAGE and Graph Transformers.
  • Ablation Studies: Removing the "mirror" component (MTGAE-ot) led to a measurable increase in error (MAE and RMSE), proving the necessity of the MTCM.
  • Generalization: On the DGraphFin financial dataset, MTGAE achieved an AUC of 0.768, demonstrating robustness beyond transportation data.

Conclusion and Future Work

The MTGAE framework effectively integrates spatial and temporal features through its mirrored convolutional approach and high-dimensional mapping. While it shows superior performance, future research will focus on adapting the model to dynamic graphs (where road structures change) and optimizing the computational complexity of the nested GCNs for real-time deployment in edge-computing sensors.

Summary updated

Abstract

Traffic time series anomaly detection has been intensively studied for years because of its potential applications in intelligent transportation. However, classical traffic anomaly detection methods often overlook the evolving dynamic associations between road network nodes, which leads to challenges in capturing the long-term temporal correlations, spatial characteristics, and abnormal node behaviors in datasets with high periodicity and trends, such as morning peak travel periods. In this paper, we propose a mirror temporal graph autoencoder (MTGAE) framework to explore anomalies and capture unseen nodes and the spatiotemporal correlation between nodes in the traffic network. Specifically, we propose the mirror temporal convolutional module to enhance feature extraction capabilities and capture hidden node-to-node features in the traffic network. Morever, we propose the graph convolutional gate recurrent unit cell (GCGRU CELL) module. This module uses Gaussian kernel functions to map data into a high-dimensional space, and enables the identification of anomalous information and potential anomalies within the complex interdependencies of the traffic network, based on prior knowledge and input data. We compared our work with several other advanced deep-learning anomaly detection models. Experimental results on the NYC dataset illustrate that our model works best compared to other models for traffic anomaly detection.
BibTeX
@inproceedings{ren2024graph,
  author = {Zhiyu Ren and Xiaojie Li and Jing Peng and Ken Chen and Qushan Tan and Xi Wu and Canghong Shi},
  title = {{Graph autoencoder with mirror temporal convolutional networks for traffic anomaly detection}},
  booktitle = {Scientific},
  year = {2024},
  doi = {10.1038/s41598-024-51374-3},
}
Figure