PaperlayerView exact revision on arXiv

Source-linked AI summary

KST-GCN: A Knowledge-Driven Spatial-Temporal Graph Convolutional Network for Traffic Forecasting

Jiawei Zhu, Xin Han, Hanhan Deng, Chao Tao, Ling Zhao, Pu Wang, Lin Tao, Haifeng Li

arXiv:2011.14992v2cs.LG

TL;DR

Traffic forecasting must capture external factors and their complex correlations with traffic, which existing methods often neglect. KST-GCN constructs a knowledge graph, learns knowledge representations, and fuses them with traffic features before a spatial-temporal graph convolutional backbone. Experiments show improved backbone performance across prediction horizons, with ablations supporting the contribution of external-factor knowledge.

  • Problem

    Existing traffic forecasting studies often neglect external factors or their complex correlations with traffic, despite these factors affecting traffic conditions.

  • Method

    KST-GCN constructs a traffic knowledge graph, learns knowledge representations, and fuses them with traffic features through a Knowledge Fusion Cell before spatial-temporal graph convolution.

  • Results

    KST-GCN models outperform their backbones across prediction horizons; at 15 minutes, KF-DCRNN reduces RMSE by 1.47% over DCRNN and KF-T-GCN reduces RMSE by 0.63% over T-GCN.

  • Takeaways & Limitations

    Knowledge representations of roads and external attributes can improve traffic forecasting backbones, with POI and weather knowledge contributing to lower prediction errors.

Abstract

from arXiv · show

While considering the spatial and temporal features of traffic, capturing the impacts of various external factors on travel is an essential step towards achieving accurate traffic forecasting. However, existing studies seldom consider external factors or neglect the effect of the complex correlations among external factors on traffic. Intuitively, knowledge graphs can naturally describe these correlations. Since knowledge graphs and traffic networks are essentially heterogeneous networks, it is challenging to integrate the information in both networks. On this background, this study presents a knowledge representation-driven traffic forecasting method based on spatial-temporal graph convolutional networks. We first construct a knowledge graph for traffic forecasting and derive knowledge representations by a knowledge representation learning method named KR-EAR. Then, we propose the Knowledge Fusion Cell (KF-Cell) to combine the knowledge and traffic features as the input of a spatial-temporal graph convolutional backbone network. Experimental results on the real-world dataset show that our strategy enhances the forecasting performances of backbones at various prediction horizons. The ablation and perturbation analysis further verify the effectiveness and robustness of the proposed method. To the best of our knowledge, this is the first study that constructs and utilizes a knowledge graph to facilitate traffic forecasting; it also offers a promising direction to integrate external information and spatial-temporal information for traffic forecasting. The source code is available at https://github.com/lehaifeng/T-GCN/tree/master/KST-GCN.

I. INTRODUCTION

Traffic forecasting must account for spatial-temporal dependencies and external factors whose semantic relationships affect traffic conditions. KST-GCN addresses this gap by constructing a traffic knowledge graph and fusing its derived knowledge with spatial-temporal graph convolutional backbones.

  • Motivation: Traffic forecasting supports urban management and helps travelers develop efficient routes.It provides prediction reports that can guide traffic decisions and route planning.
  • Existing forecasting models: Traffic states depend on spatial and temporal correlations across upstream, downstream, and adjacent roads.Existing approaches include recurrent networks for temporal dependence and graph-based models for road-network spatial dependence.
  • Research gap: External factors such as weather, stations, emergencies, holidays, and nearby POIs affect traffic through direct or indirect relationships.Road sections can respond differently to the same factor because their surrounding attributes differ.
  • Research gap: Existing studies often include external factors without modeling their interrelationships with traffic information.The paper identifies semantic correlation among multisource data as central to improving traffic-flow prediction.
  • Proposed approach: KST-GCN constructs a traffic knowledge graph, derives knowledge representations, and incorporates them through a Knowledge Fusion Cell.The fused knowledge is supplied to backbone spatial-temporal graph convolutional networks.
  • Evaluation: Across prediction horizons, KF-Cell improves various backbone models, while ablations examine semantic relationships and dynamic versus static external factors.These evaluations are conducted on a real-world dataset.

A. Traffic Flow Forecasting

Traffic forecasting methods progressed from statistical and neural models toward graph-based approaches, but integrating external-factor information remains challenging. KST-GCN addresses this by combining knowledge representations with traffic-network features before spatial-temporal graph convolution.

  • Conventional approaches: Conventional traffic forecasting uses historical averages, time-series models, and Kalman filtering to predict traffic conditions.These approaches rely primarily on statistical analysis of traffic data.
  • Learning-based approaches: Machine-learning and deep-learning models learn nonlinear patterns from large quantities of historical traffic data.The passage distinguishes conventional machine-learning approaches from deep-learning approaches.
  • Spatial-temporal modeling: Graph convolutional networks model road-network topology, whereas recurrent models capture temporal tendencies and periodicity.The related models may separately emphasize spatial structure or temporal behavior.
  • External information: The main challenge is integrating external information such as weather, stations, bus stops, and POIs into traffic prediction models.Prior multisource methods are introduced as attempts to address this integration problem.
  • Knowledge graphs: Knowledge graphs offer a way to represent heterogeneous multisource relations through structural and relational information.Their use follows earlier heterogeneous-network approaches that can require carefully designed metapaths.
  • KST-GCN framework: KST-GCN constructs a traffic knowledge graph, derives correlation-aware embeddings, and fuses them with traffic features before spatial-temporal graph convolution.The framework function uses road topology A, node features X, and the traffic knowledge graph CKG.

B. Knowledge Graph Representation Learning

The paper represents traffic information and external factors in a heterogeneous knowledge graph, then uses KR-EAR to learn separate representations for entities, relations, and attributes. This representation distinguishes attribute and relational information to capture semantic structure and correlations between road sections and external factors.

  • Knowledge graph construction: Knowledge graphs represent heterogeneous nodes and multi-relational information as semantic triples linking entities through relations.The paper uses this structure to preserve semantic information and network structure across data sources.
  • KR-EAR representation: KR-EAR separates attribute triples from relation triples instead of modeling all triples with a single entity-relation representation.This distinction reflects the different meanings of entity relationships and entity properties.
  • KR-EAR representation: The entity-attribute-relationship representation is selected because road-network relationships with attribute factors include one-to-many and many-to-one patterns.The paper states that distinguishing attribute and relational information is therefore more suitable for this setting.
  • Knowledge graph construction: The traffic knowledge graph contains road-adjacency triples, road-attribute triples, and attribute co-occurrence relations.Road-adjacency triples connect road sections; attribute triples associate roads with attribute values; co-occurrence relations encode relationships between attribute classes.
  • KR-EAR representation: KR-EAR learns embeddings for entities, relations, and attributes by maximizing the joint probability of relation and attribute triples.Relation triples use a TransR-based conditional probability, while attribute triples use a classification model to encode entity-attribute correlations.

C. KF-Cell

The KF-Cell introduces knowledge embeddings and traffic features into spatial-temporal graph convolutional backbones. It combines static and dynamic external-factor representations, applies graph convolution for spatial dependence, and uses recurrent processing for temporal dependence before prediction.

  • KF-Cell design: KF-Cell takes knowledge embeddings and road-section traffic features as inputs, enabling external knowledge to enter spatial-temporal graph convolutional networks.The cell is designed to adapt to traffic forecasting models based on Graph Neural Networks and Recurrent Neural Networks.
  • Knowledge fusion: Static and dynamic external factors are represented separately by road-section embeddings e_s and e_d produced after KR-EAR.Linear transformations and bias constants are applied before the resulting features are combined with traffic features.
  • Spatial-temporal backbone: The updated road-section features X′_t and adjacency matrix A are input to GCN layers to model spatial dependence in the traffic network.The GCN captures topological relations and features for each road section, including influences from connected road sections.
  • Spatial-temporal backbone: The recurrent component uses a GRU with reset and update gates to capture temporal dependence after graph-based spatial processing.The reset gate combines memory with current information, while the update gate selects or forgets memories.
  • Prediction: The recurrent state h_t is passed to a fully connected layer to generate predicted future speed, with training minimizing prediction error against observed speed.This defines the final forecasting stage of the knowledge-driven model.

A. Data Description

The study uses Shenzhen data combining traffic, road-network, weather, and POI information, then represents road and external-factor relationships as knowledge-graph triples.

  • Data Description: The experiments use one Luohu District, Shenzhen dataset because jointly collecting aligned road, traffic, POI, and weather data is difficult.The authors note that the setup can be validated in other cities when equivalent data are available.
  • Data Description: The dataset contains taxi tracks, road-network data, weather data, and street-level POIs collected from January 1 to January 31, 2015.It covers 156 road sections and nine POI categories.
  • Data Description: Road sections, POI categories, POI counts, time, and weather conditions are encoded as attribute and relational triples.Examples include road-section restaurants or schools and road-section weather conditions at a given moment.

B. Experimental Settings

The experiments test knowledge-fused variants of DCRNN and T-GCN, selecting embedding and hidden-unit settings through hyperparameter experiments.

  • Experimental Settings: The method is evaluated with KF-DCRNN and KF-T-GCN, knowledge-fused variants of DCRNN and T-GCN.The KF-Cell is designed to adapt to traffic-forecasting models based on graph and recurrent neural networks.
  • Experimental Settings: The experiments use a 0.001 learning rate and allocate 80% of the dataset for training.Embedding dimension and hidden units are identified as important hyperparameters.
  • Experimental Settings: KF-T-GCN performs best with 128 hidden units under the reported hyperparameter comparison.The comparison evaluates hidden-unit counts from 16 to 256.

1) Prediction Accuracy

Knowledge fusion improves both backbones over their unfused versions across prediction horizons, with larger reported gains at longer horizons.

  • 1) Prediction Accuracy: At a 15-minute horizon, KF-DCRNN reduces RMSE by 1.47% over DCRNN, while KF-T-GCN reduces RMSE by 0.63% over T-GCN.The KST-GCN variants also outperform traditional and single-dependency deep-learning baselines in the reported comparison.
  • 1) Prediction Accuracy: Across prediction horizons, KF-DCRNN’s RMSE reduction grows from 1.50% to 2.85%, while its Accuracy increment grows from 0.59% to 1.15%.The underlying models all deteriorate as prediction horizons increase, but KF-DCRNN remains better than DCRNN.
  • 1) Prediction Accuracy: Across prediction horizons, KF-T-GCN’s RMSE reduction grows from 0.62% to 4.36%, while its Accuracy increment grows from 0.57% to 0.97%.The reported pattern supports superiority over T-GCN for both short-term and long-term prediction.
  • 1) Prediction Accuracy: KF-T-GCN has lower RMSE than AST-GCN for long-term prediction and shows milder fluctuation, although its RMSE is slightly higher at 15 and 30 minutes.The comparison evaluates knowledge representation against direct external-factor concatenation.

4) Ablation Study

Ablation results show that combining POI and weather knowledge improves both backbones, while perturbation experiments report robustness to added noise.

  • 4) Ablation Study: Combining POI and weather knowledge lowers KF-T-GCN’s prediction error by 0.62% relative to T-GCN.Using weather alone lowers RMSE by 0.28%, and using POI alone lowers RMSE by 0.32%.
  • 4) Ablation Study: Combining POI and weather knowledge lowers KF-DCRNN’s prediction error by 1.47% relative to DCRNN.Using weather alone lowers RMSE by 1.00%, and using POI alone lowers RMSE by 1.06%.
  • 4) Ablation Study: The ablation results report stronger performance for POI knowledge than weather knowledge, indicating a more pronounced road–POI semantic correlation for traffic prediction.The authors conclude that fusing roads with knowledge improves prediction accuracy to some extent.
  • Robustness Analysis: Adding Gaussian or Poisson noise to the urban traffic data does not significantly affect the KST-GCNs.The reported conclusion is that the KST-GCNs are robust to possible data noise.

D. Model Interpretation

KF-T-GCN captures traffic trends across all prediction horizons, with closer agreement at short horizons than long horizons. Deviations at turning points suggest that weather, POIs, and traffic contingencies jointly affect peak variations.

  • KF-T-GCN captures the changing tendency of traffic data across all evaluated prediction horizons.The visualized horizons are 15, 30, 45, and 60 minutes.
  • Short-range predictions are closer to ground truth than long-range predictions, with 15-minute results outperforming 60-minute results.The authors attribute this pattern to more complex spatial-temporal dependencies in long-term prediction.
  • KF-T-GCN shows significant deviations at turning points in the traffic series.The authors relate peak variations to weather, POIs, and traffic contingencies acting in combination.
  • KST-GCNs improve backbone forecasting performance across various prediction horizons and show effectiveness and robustness in ablation and perturbation analyses.The models use knowledge representations and a Knowledge Fusion Cell to incorporate external-factor correlations into spatial-temporal graph convolutional networks.
Loading 2011.14992v2…