PaperlayerView exact revision on arXiv

Source-linked AI summary

Hybrid Spatio-Temporal Graph Convolutional Network: Improving Traffic Prediction with Navigation Data

Rui Dai, Shenkun Xu, Qian Gu, Chenguang Ji, Kaikui Liu

arXiv:2006.12715v1cs.LGstat.ML

TL;DR

Stochastic events make traffic non-stationary, limiting travel-time forecasting when contextual information is absent. H-STGCN converts navigation-derived intended traffic volume into travel time and combines it with travel-time signals using graph convolution. On real-world datasets, it reportedly improves over benchmark methods, especially for non-recurring congestion.

  • Problem

    Stochastic events make travel demand difficult to predict, and missing contextual information renders road traffic non-stationary, limiting travel-time forecasting.

    Source: §1 · p. 1

  • Method

    H-STGCN converts navigation-derived intended traffic volume into equivalent travel time, combines it with travel-time data, and applies temporal and graph convolution using compound adjacency.

    Source: §1 · pp. 1–2

  • Results

    H-STGCN achieves remarkable improvement over benchmark methods on real-world datasets, especially for predicting non-recurring congestion.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    The architecture offers a formalism for embedding physics knowledge in data-driven models that can be applied to general spatio-temporal forecasting tasks.

    Source: §6 · p. 9

    Ideal future volume assumes only navigation-service users are considered and each follows the exact planned path at a speed consistent with ETA.

    Source: §2 · pp. 2–3

Abstract

from arXiv · show

Traffic forecasting has recently attracted increasing interest due to the popularity of online navigation services, ridesharing and smart city projects. Owing to the non-stationary nature of road traffic, forecasting accuracy is fundamentally limited by the lack of contextual information. To address this issue, we propose the Hybrid Spatio-Temporal Graph Convolutional Network (H-STGCN), which is able to "deduce" future travel time by exploiting the data of upcoming traffic volume. Specifically, we propose an algorithm to acquire the upcoming traffic volume from an online navigation engine. Taking advantage of the piecewise-linear flow-density relationship, a novel transformer structure converts the upcoming volume into its equivalent in travel time. We combine this signal with the commonly-utilized travel-time signal, and then apply graph convolution to capture the spatial dependency. Particularly, we construct a compound adjacency matrix which reflects the innate traffic proximity. We conduct extensive experiments on real-world datasets. The results show that H-STGCN remarkably outperforms state-of-the-art methods in various metrics, especially for the prediction of non-recurring congestion.

1 INTRODUCTION

Traffic prediction is difficult when stochastic events make road traffic non-stationary and contextual information is missing. H-STGCN uses intended traffic volume from navigation routes alongside travel-time data, with a flow-to-time transformer and compound graph adjacency; experiments report improvements over state-of-the-art benchmarks.

1 INTRODUCTION

  • Traffic prediction supports applications including time-dependent routing, intelligent traffic-light control, and proactive traffic management.Source: §1 · p. 1
  • Stochastic events make travel-time forecasting difficult because missing contextual information renders road traffic non-stationary; event-related features have limited efficacy around hotspots.Earlier data-driven approaches using travel time as the main input performed suboptimally.Source: §1 · p. 1
  • Aggregated planned routes provide intended traffic volume that reflects live travel demand and offers clues to future travel time.The paper describes this navigation-derived signal as more detailed than numerous event-level features.Source: §1 · p. 1
  • A two-stage flow-to-time transformer captures shared flow-curve geometry and segment-specific characteristics when converting volume into travel time.The design draws on the triangular flow-density relationship, whose peak capacity varies by road segment.Source: §1 · p. 2, Figure 2 · p. 2
  • The compound adjacency matrix combines spatial attenuation with travel-time covariance to better reflect innate traffic proximity.This addresses cases where congestion on a major thoroughfare may not propagate to a contiguous private road.Source: §1 · p. 2
  • H-STGCN combines intended traffic volume and travel-time signals with temporal and graph convolution, and outperforms state-of-the-art benchmarks on real-world datasets.The model converts intended volume into equivalent travel time before applying shared temporal convolution and graph convolution with compound adjacency.Source: §1 · p. 2

2 PRELIMINARIES

The paper defines traffic inputs and the forecasting task over a regional road network divided into directional segments and five-minute intervals. It also defines ideal future volume under assumptions about navigation-service users, planned routes, and travel speed.

2 PRELIMINARIES

  • The network is divided into directional road segments at intersections, and data are represented as a tensor with segment, time, and input-feature dimensions.Time is split into five-minute intervals; the training and test ranges are denoted [0, S_train) and [S_train, S_train + S_test).Source: §2 · p. 2
  • Travel time is average traversing time per unit length on a segment, while traffic volume counts vehicles entering that segment during a time slot.Source: §2 · p. 2
  • Ideal future volume corresponds to future segment volume under assumptions that only current navigation-service users count and each follows the exact planned path at ETA-consistent speed.Source: §2 · pp. 2–3
  • The forecasting task uses previous traffic features to predict network-wide travel time for the next F slots.P denotes the input-series length, and F the forecast horizon.Source: §2 · p. 3

3 METHODOLOGY

H-STGCN combines intended future traffic volume with travel time, translating volume into a travel-time signal before temporal and graph convolutions produce forecasts. Its design pairs shared temporal processing with a compound adjacency matrix that accounts for traffic proximity.

3.1 Overall Architecture

  • H-STGCN transforms ideal future volume into travel time, processes it alongside observed travel time, and applies temporal and graph convolutions before a final fully connected layer outputs forecasts.The graph convolution processes the concatenated temporal signals to capture interactions between traffic volume and travel time.Source: §3.1 · p. 3

3.2 Model Input and Data Processing

  • Amap navigation records provide planned routes and estimated arrival times, which are aggregated into ideal future volume; historical averages and travel-time data also enter the model.The model uses ideal future volume within the prediction window and its corresponding historical average as input.Source: §3.2 · pp. 3–4

3.3 Domain Transformer

  • The domain transformer converts ideal future volume into equivalent travel time using shared and segment-specific 1 × 1 convolutions.The shared layer captures a universal triangular-shaped mapping, while segment-specific parameters provide capacity for segment-level features.Source: §3.3 · p. 4

3.4 Graph Convolution with Compound Adjacency Matrix

  • The compound adjacency matrix combines travel-time covariance with distance-decaying Dijkstra proximity, reflecting nonuniform congestion propagation while limiting long-range correlations.The covariance term is adjusted to remove the correlation floor associated with common free-flow periods.Source: §3.4 · p. 5

3.5 Temporal Gated Convolution

  • A shared gated 1D convolution extracts temporal features separately for each road segment, with a gated linear unit transforming its convolution outputs.The gate combines equal-sized A and B tensors through element-wise multiplication of A by the sigmoid of B.Source: §3.5 · p. 5

3.6 Connection to STGCN

  • Removing the volume-processing branch and covariance term reduces H-STGCN to an STGCN model with a single spatio-temporal convolution block.The referenced STGCN alternates spatial graph convolution and temporal gated convolution layers.Source: §3.6 · pp. 5–6

3.7 Model Training

  • Training adds Gaussian noise to low-volume channels to mitigate overfitting and uses L1 loss for multistep traffic forecasting.The loss compares model output with ground truth.Source: §3.7 · p. 6

4 EXPERIMENTS

Experiments on two Beijing road networks compare H-STGCN with traditional and neural-network baselines, using full, congested, and non-recurring-congestion test sets. H-STGCN outperforms the benchmarks across metrics, while ablations examine the compound adjacency matrix and future-volume feature.

4.1 Datasets

  • The datasets cover 715 segments around Beijing’s West 3rd Ring Road and 2,907 around its East 5th Ring Road, with eight weeks for training and two for testing.Data span 06:00–22:00 from December 24, 2018 to April 21, 2019, with holidays removed.Source: §4.1 · p. 6

4.2 Compared Methods

  • Comparisons include historical and machine-learning baselines, MLP, Seq2Seq, STGCN, and variants that change the adjacency matrix or set input volume to one.The STGCN baseline uses a single ST-Conv block; STGCN (Im) substitutes the compound adjacency matrix.Source: §4.2 · pp. 6–7

4.3 Implementation Details

  • Models use six previous time slots as input to predict the next hour, and evaluation uses MAE, MAPE, and RMSE across full, congested, and non-recurring-congested test sets.Congested test subsets focus on high-volume segments; NRC periods are extended by an hour in each direction to include formation and dissipation.Source: §4.3 · p. 7, §4.4 · p. 7
  • The domain transformer uses 16 filters in its shared and segmentwise 1 × 1 convolutions, while graph convolution uses 64 filters.The final fully connected layer outputs 12 values for the forecasting period.Source: §4.3 · p. 7

4.5 Performance Comparison

  • The compound adjacency variant improves on STGCN: it lowers MAE and MAPE on W3-715 and all three metrics on E5-2907.The comparison is between STGCN and STGCN (Im), which uses the compound adjacency matrix.Source: §4.5 · p. 7, Table 3 · p. 8
  • On the NRC test set, the future-volume feature becomes the dominant contributor to error reduction as the forecasting horizon lengthens.An example describes H-STGCN forecasting congestion 30 minutes ahead using ideal future volume, while H-STGCN (1) shows a time lag.Source: Figure 8 · p. 7, §4.5 · pp. 7–8
  • H-STGCN significantly outperforms the benchmarks in all metrics, especially for non-recurring congestion.Table 3 compares baselines on the full test set and the congested (C) and non-recurring-congested (NRC) test sets.Source: §4.5 · p. 7, Table 3 · p. 8

4.6 Model Scalability

  • Inference takes less than 100 ms on the reported networks; the authors propose partitioning city-wide networks into sub-networks of at most a few thousand segments.Partitioning minimizes congested boundary links, with a separate model trained and deployed for each sub-network.Source: §4.6 · p. 8

5 RELATED WORK

Prior traffic-prediction work includes theory-driven simulation and data-driven models, with deep learning extending spatial and temporal modeling. Auxiliary context has been explored for non-recurring congestion, while forecasting spatial resolution remains a concern for critical real-world applications.

Approaches to traffic prediction

  • Theory-driven methods simulate traffic dynamics, while data-driven work ranges from shallow models to deep learning approaches.The paper describes shallow models as limited in representation power and notes deep-learning improvements over traditional benchmarks.Source: §5 · p. 8

Deep-learning approaches

  • Deep-learning methods use graph convolution for spatial dependency and encoder-decoder or convolutional architectures for temporal patterns.The related work also notes non-local correlation mechanisms for global spatial correlations.Source: §5 · p. 8

Non-recurring congestion

  • For non-recurring congestion, prior studies use auxiliary context such as weather, tweets, road structure, points of interest, or crowd-map queries, but spatial resolution remains insufficient for critical real-world applications.Source: §5 · p. 8

6 CONCLUSION

The conclusion presents H-STGCN as a travel-time forecasting architecture that uses intended traffic-volume data and reports improved performance over benchmark methods, especially for non-recurring congestion. It also frames the architecture as a way to embed physics knowledge in data-driven spatio-temporal forecasting.

  • H-STGCN is a deep travel-time forecasting architecture featuring intended-traffic-volume data.Source: §6 · p. 8
  • Figure 9 depicts a non-recurring congestion prediction case on a freeway segment on April 16, 2018.Its caption defines GT as ground truth, HA as historical average, and νi,t,f as ideal future traffic volume for the corresponding future time slot.Source: Figure 9 · p. 9
  • H-STGCN achieves remarkable improvement over benchmark methods, especially for predicting non-recurring congestion.Source: §6 · p. 9
  • The conclusion presents the architecture as a formalism for embedding physics knowledge in data-driven models for general spatio-temporal forecasting tasks.Source: §6 · p. 9
Loading 2006.12715v1…