PaperlayerView exact revision on arXiv

Source-linked AI summary

DeepTransport: Learning Spatial-Temporal Dependency for Traffic Condition Forecasting

Xingyi Cheng, Ruiqing Zhang, Jie Zhou, Wei Xu

arXiv:1709.09585v4cs.AI

TL;DR

Traffic prediction methods often use only the target location and nearby regions, while current traffic conditions alone are insufficient for effective route planning. DeepTransport models temporal and spatial dependencies using a deep architecture and attention, and the authors report that it significantly outperforms prior statistical and deep-learning methods.

  • Problem

    Existing traffic forecasting approaches use target-location observations and adjacent regions but provide limited accuracy when road topology and wider surrounding locations are not accounted for.

    Source: Introduction · p. 2

  • Method

    DeepTransport uses a deep architecture for temporal and dynamic spatial modeling, with attention to quantify spatial-temporal relationships.

    Source: Introduction · p. 2

  • Results

    DeepTransport significantly outperforms previous statistical and deep-learning methods for traffic forecasting.

    Source: Introduction · p. 2, Conclusions · p. 8

  • Takeaways & Limitations

    The experiments support the importance of using road temporal and spatial information in traffic-condition forecasting.

    Source: Conclusions · p. 8

Abstract

from arXiv · show

Predicting traffic conditions has been recently explored as a way to relieve traffic congestion. Several pioneering approaches have been proposed based on traffic observations of the target location as well as its adjacent regions, but they obtain somewhat limited accuracy due to a lack of mining road topology. To address the effect attenuation problem, we suggest taking into account the traffic of surrounding locations(wider than the adjacent range). We propose an end-to-end framework called DeepTransport, in which Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) are utilized to obtain spatial-temporal traffic information within a transport network topology. In addition, an attention mechanism is introduced to align spatial and temporal information. Moreover, we constructed and released a real-world large traffic condition dataset with a 5-minute resolution. Our experiments on this dataset demonstrate our method captures the complex relationship in the temporal and spatial domains. It significantly outperforms traditional statistical methods and a state-of-the-art deep learning method.

Introduction

Traffic forecasting is needed because current conditions alone are insufficient for effective route planning, while existing approaches often underuse road topology. DeepTransport models spatial-temporal dependencies with neural networks and attention, and the reported experiments outperform earlier statistical and deep-learning methods.

Introduction

  • Earlier traffic prediction approaches include classical statistical time-series analysis and machine-learning methods, with ARIMA among the foundational statistical methods.The cited passage names ARIMA and its variations alongside multivariate approaches using temporal and spatial features.Source: Introduction · p. 1
  • Existing approaches either focus on time series or use little spatial information, and narrow spatial ranges can degrade forecasting accuracy, especially at long horizons.Earlier deep-learning methods concatenated observations from different locations without modeling their spatial relationships.Source: Introduction · p. 2
  • DeepTransport combines CNNs to preserve the spatial “width” structure with LSTM recurrent networks for the “depth” sequence of surrounding locations.The model uses attention to represent relationships between spatial and temporal information.Source: Introduction · p. 2
  • The proposed architecture is designed to model temporal and dynamic spatial information for traffic-condition forecasting.The paper presents attention as a way to quantify spatial-temporal relationships.Source: Introduction · p. 2
  • DeepTransport significantly outperforms existing deep-learning and time-series forecasting methods.The paper also reports releasing a large traffic dataset with topological networks and temporal conditions.Source: Introduction · p. 2
  • The authors release a real large traffic dataset with millions of records, topological networks, and temporal traffic conditions.It was developed for the ASC Student Supercomputer Challenge 2017 on PaddlePaddle.Source: Introduction · p. 2

Preliminary

The preliminary section represents traffic networks as directed graphs and defines the spatial and temporal observations used to forecast conditions at a target road location. It introduces order slots and perceptive radius to specify the surrounding locations included.

Preliminary

  • Traffic condition colors encode states from fluency to extreme congestion, and the numeric label identifies a road section.The figure caption lists fluency, slow, congestion, and extreme congestion as displayed states.Source: Figure 2 · p. 2

Common Notations and Definition

  • The network is a directed graph whose vertices can represent road stretches; each location has one of five traffic-condition states.The states include not-released, fluency, slow, congestion, and extreme congestion.Source: Common Notations and Definition · p. 2
  • Each vertex has time-varying traffic observations aggregated at five-minute intervals and time-invariant road features such as length and speed limit.The observations are paired with static properties, including road input and output degrees.Source: Common Notations and Definition · p. 2
  • An order slot groups vertices by their edge distance from a target, with directly linked vertices first and their neighbors in the next order.For a directed network, the paper distinguishes upstream and downstream flow and illustrates first- and second-order slots around L4.Source: Common Notations and Definition · pp. 2–3, Figure 3 · p. 3
  • The perceptive radius is the maximum order considered, and forecasting predicts a future traffic value from historical observations within that radius.The problem definition specifies a history of length p and surrounding locations through order r.Source: Common Notations and Definition · p. 3

Model

DeepTransport combines upstream and downstream traffic-flow modules with a target-location module and multi-task output layers. CNNs and recurrent networks represent road topology, while attention aligns representations from different-order slots.

Model

  • The architecture has upstream-flow, target-location, downstream-flow, and training-cost modules, with the target representation concatenated with both flow representations before the cost layer.The example architecture shows three slots on each flow side and four cost layers.Source: Model · p. 3, Figure 4 · p. 4, Slot Attention · p. 5

Spatial-temporal Relation Construction

  • CNNs encode width-wise spatial information, and separate recurrent networks represent upstream and downstream paths toward or away from the target.The convolution uses multiple encoders and shared weights to limit path-related dimensionality and computation.Source: Convolutional Layer · p. 3, Recurrent Layer · p. 3
  • LSTM outputs are pooled into order-specific slot embeddings, which target-conditioned attention weights combine into upstream and downstream representations.The attention weights are computed from each slot's relevance to the target-location hidden representation.Source: Slot Attention · pp. 4–5

Top Layers with Multi-task Learning

  • The model shares most parameters across prediction-time tasks, while keeping attention weights and output-layer affine transformations task-specific.The paper treats different future time points as different tasks.Source: Top Layers with Multi-task Learning · p. 5

Experiments

The experiments use MapBJ, a Beijing traffic dataset collected at five-minute intervals, and evaluate categorical predictions with quadratic weighted Cohen’s Kappa.

Dataset

  • MapBJ contains about 349 Beijing locations, five-minute traffic conditions, road-topology graphs, and road limit-speed information.The data were collected from March 2016 to June; the first two months were used for training and the remaining half month for testing.Source: Dataset · p. 5
  • The dataset labels traffic conditions as fluency, slow, congestion, or extreme congestion, and discretizes road limit speed into four classes.Source: Dataset · p. 5

Evaluation

  • Performance is evaluated with quadratic weighted Cohen’s Kappa, using observed, expected, and rating-weight matrices for ground-truth and predicted ratings.The metric ranges typically from 0 for random agreement to 1 for complete agreement.Source: Evaluation · p. 5

Implementation Details

Implementation uses square-error training and Adam, with historical-measurement length and perceptive radius controlling temporal and spatial scope. Hyperparameter analysis measures dependence between locations and examines validation performance across radius and history length.

Implementation Details

  • Square error was used to fit future traffic conditions because softmax cross-entropy ignores the magnitude of the four condition ratings.Source: Implementation Details · p. 5
  • Prediction results are ranked in ascending order and assigned to four traffic classes: the first 88.2% as fluency, 88.2%-96.7% as slower, 96.7%-99.5% as congestion, and 99.5%-100% as extreme congestion.Source: Implementation Details · p. 5
  • The model was optimized with Adam using batch size 1100 and trained until convergence; historical length p and perceptive radius r control temporal and spatial scope.Observations were represented as 32-dimensional continuous vectors.Source: Implementation Details · p. 5

Choosing Hyperparamerters

  • NMI between a location and its rth-order neighbors decreases from 0.116 at r=1 to 0.034 at r=5, then hardly changes for r > 5.The reported values for r=2, 3, and 4 are 0.052, 0.038, and 0.035, respectively.Source: Choosing Hyperparamerters · p. 6
  • The study sets p to {3, 6, 12, 18}, corresponding to 15, 30, 60, and 90 minutes, and r to {1, 2, 3, 4, 5}.Source: Choosing Hyperparamerters · p. 6

Effects of Hyperparameters

Validation results show that perceptive radius and historical-measurement length interact with prediction horizon: wider radii matter more for longer horizons, while longer histories do not improve performance monotonically.

Effects of Hyperparameters

  • Figure 5 reports averaged quadratic weighted kappa over perceptive radius r and historical-measurement length p on validation data.Source: Effects of Hyperparameters · p. 6
  • A closer prediction horizon performs better, and increasing r has a larger effect at longer horizons, including a steeper r=1-to-5 slope at 60 minutes than at 15 minutes.Source: Effects of Hyperparameters · p. 6
  • For a 60-minute estimate, validation performance does not rise monotonically with history length p; the same pattern occurs at other prediction horizons.The passage attributes this pattern to more parameters with larger p, leading to overfitting.Source: Effects of Hyperparameters · p. 6

Comparison with Other Methods

DeepTransport is compared with statistical and neural baselines, including spatial-input models. The reported comparison favors models using spatial information, with DeepTransport also accounting for transport-network structure.

Comparison with Other Methods

  • The comparison includes RW, ARIMA, SAEs, and DeepTransport; the passage also describes an FNN baseline.FNN is a single-hidden-layer network with 32 hidden neurons and four horizon outputs.Source: Comparison with Other Methods · p. 6
  • The reported DeepTransport configurations are R1P12 (r = 1, p = 12) and R5P12 (r = 5, p = 12).Source: Comparison with Other Methods · p. 6
  • Models using spatial information, including SAEs and DeepTransport, significantly outperform RW, ARIMA, and FNN, especially at longer prediction horizons.The authors attribute DeepTransport's higher performance relative to SAEs to its consideration of transport-network structure.Source: Comparison with Other Methods · p. 6
  • Unlike SAEs' fully connected representation, DeepTransport considers transport-network structure when modeling traffic.The authors describe SAEs as assuming any pair of locations is directly connected, thereby neglecting network topology.Source: Comparison with Other Methods · p. 6

Slot Attention Weights

DeepTransport's attention weights show different spatial-order patterns for upstream and downstream dependencies. Downstream attention shifts toward higher-order slots at longer prediction times, while upstream attention remains strongest at lower orders.

Slot Attention Weights

  • Averaging attention weights across target locations lets the model show each slot's influence on the target location.Source: Slot Attention Weights · pp. 6–7
  • As prediction time increases, downstream attention shifts from lower-order to higher-order slots, whereas upstream attention favors lower-order slots.Source: Slot Attention Weights · p. 7, Figure 6 · p. 7
  • The authors relate downstream higher-order attention to accumulated congestion affecting the target, while dispersed upstream flow makes nearby locations more influential.They describe the upstream first-order slot as having more impact for any future time.Source: Slot Attention Weights · p. 7

Case Study

The case study finds DeepTransport's advantage most clearly during rush hours and just after traffic peaks. During flat periods, especially early morning, the models show little difference.

Case Study

  • During flat periods, especially early morning, there is almost no difference between models.The authors attribute this to roads being nearly free-flowing.Source: Case Study · p. 7
  • During rush hours, DeepTransport has lower errors than other models at 15 minutes, with a more pronounced advantage at 60 minutes.Source: Case Study · p. 7
  • After traffic peaks, DeepTransport predicts better over the periods when traffic conditions may be easing.Source: Case Study · p. 7

Related Works

Related work spans statistical and data-driven traffic prediction, deep-learning traffic models, and spatial-temporal models for citywide crowds. The authors distinguish their traffic-network setting from approaches built on uniform city grids.

Related Works

  • Statistical traffic models such as ARIMA have a long history, but their assumptions may fail on complex, highly nonlinear data.Source: Related Works · p. 7
  • Deep traffic-prediction work includes DBNs, stacked autoencoders, and FNNs that take concatenated observations from multiple locations as input.Source: Related Works · p. 7
  • DeepST and ST-ResNet use convolutional or residual networks to forecast citywide crowd flows on longitude-latitude grids.Source: Related Works · p. 7
  • The authors do not adopt these grid-based approaches because MapBJ represents traffic as road sections and partitions roads by speed-limit level rather than equal length.Source: Related Works · pp. 7–8

Conclusions

DeepTransport combines spatial-temporal modeling with attention, and the authors report that it outperforms prior statistical and deep-learning methods for traffic forecasting.

  • DeepTransport combines CNN and RNN sequential models with attention to capture spatial-temporal information and quantify dependency relationships.Source: Conclusions · p. 8
  • The authors released a real-world traffic-condition dataset containing millions of recordings.Source: Conclusions · p. 8
  • DeepTransport significantly outperformed previous statistical and deep-learning methods for traffic forecasting.Source: Conclusions · p. 8
Loading 1709.09585v4…