[Paper Review] STG2Seq: Spatial-temporal Graph to Sequence Model for Multi-step Passenger Demand Forecasting

Lei Bai, Lina Yao+3 more|arXiv (Cornell University)|May 24, 2019

Traffic Prediction and Management Techniques18 references21 citations

TL;DR

This paper proposes STG2Seq, a graph-based sequence-to-sequence model that leverages hierarchical graph convolutional networks to capture complex spatial-temporal dependencies in multi-step passenger demand forecasting. By employing a long-term encoder, short-term encoder, and attention-based output module, STG2Seq outperforms state-of-the-art methods by up to 4.9% in RMSE and 4.5% in MAPE across three real-world datasets, demonstrating superior accuracy and reduced error accumulation in long-horizon predictions.

ABSTRACT

Multi-step passenger demand forecasting is a crucial task in on-demand vehicle sharing services. However, predicting passenger demand over multiple time horizons is generally challenging due to the nonlinear and dynamic spatial-temporal dependencies. In this work, we propose to model multi-step citywide passenger demand prediction based on a graph and use a hierarchical graph convolutional structure to capture both spatial and temporal correlations simultaneously. Our model consists of three parts: 1) a long-term encoder to encode historical passenger demands; 2) a short-term encoder to derive the next-step prediction for generating multi-step prediction; 3) an attention-based output module to model the dynamic temporal and channel-wise information. Experiments on three real-world datasets show that our model consistently outperforms many baseline methods and state-of-the-art models.

Motivation & Objective

• To address the challenge of multi-step passenger demand forecasting in on-demand mobility services, where inaccurate predictions lead to inefficient vehicle dispatching.
• To overcome limitations of existing RNN- and CNN-based models that struggle with non-Euclidean spatial correlations and long-term temporal dependency.
• To reduce error accumulation in multi-step forecasting by replacing RNN decoders with a dual-encoder architecture.
• To model dynamic temporal correlations by incorporating learnable attention mechanisms that adaptively weight historical time steps.
• To ensure generalization across irregular spatial partitions, such as road network-based city regions, without requiring grid-based spatial discretization.

Proposed method

• The model formulates citywide passenger demand as a spatial-temporal graph, where each region is a node and spatial relationships are modeled via graph convolutional layers.
• A Gated Graph Convolutional Module (GGCM) with multiple GCN layers captures both spatial and temporal correlations hierarchically, reducing path length for long-range dependencies.
• The long-term encoder processes historical demand sequences to capture long-range temporal patterns, while the short-term encoder generates next-step predictions for multi-step forecasting.
• An attention-based output module dynamically models temporal and channel-wise dependencies, allowing the model to weigh past time steps adaptively based on their relevance to the target step.
• The model uses teacher forcing during training to stabilize multi-step prediction generation, improving generalization and reducing error propagation.
• The architecture is flexible and generalizable to irregular spatial partitions, such as road network-based regions, without requiring grid-based spatial discretization.

Experimental results

Research questions

• RQ1 Can a graph-based sequence-to-sequence model effectively capture non-Euclidean spatial correlations in passenger demand forecasting, especially when remote regions with similar characteristics are more influential than neighboring ones?
• RQ2 How does replacing RNN-based decoders with a dual-encoder architecture affect error accumulation in multi-step passenger demand prediction?
• RQ3 To what extent can an attention mechanism that models dynamic temporal correlations improve prediction accuracy compared to static or fixed-weighted historical aggregation?
• RQ4 How does the model perform when the city is partitioned into irregular regions based on road networks, rather than regular grids?
• RQ5 Does the proposed STG2Seq model consistently outperform state-of-the-art methods across diverse real-world datasets with varying spatial and temporal dynamics?

Key findings

• STG2Seq achieves a 4.9% relative improvement in RMSE and a 4.5% relative improvement in MAPE over the best state-of-the-art method on the DidiSY dataset, demonstrating significant performance gains.
• On the NYC and DidiSY datasets, STG2Seq reduces RMSE by 2.6% and 4.9% respectively, and MAE by 0.4% and 2.6%, showing consistent superiority across datasets.
• In multi-step prediction, STG2Seq maintains low error growth across all three time steps, while ConvLSTM and DCRNN show rapid deterioration, confirming reduced error accumulation.
• Ablation studies show that removing the short-term encoder reduces performance to single-step prediction, and replacing attention modules or the gating mechanism with ReLU or CNN layers leads to measurable performance drops.
• On irregularly partitioned regions (Shenyang based on road networks), STG2Seq achieves RMSE of 3.244 and MAE of 2.136, outperforming ARIMA, SARIMA, MLP, XGBoost, DCRNN, and STGCN.
• The model’s performance remains robust across different spatial partitioning schemes, confirming its generalization capability beyond regular grid-based city representations.