Source-linked AI summary
STG2Seq: Spatial-temporal Graph to Sequence Model for Multi-step Passenger Demand Forecasting
Lei Bai, Lina Yao, Salil. S Kanhere, Xianzhi Wang, Quan. Z Sheng
TL;DR
Accurate multi-step passenger demand forecasting is difficult because demand depends on nonlinear, dynamic spatial-temporal relationships. STG2Seq models citywide demand as a graph using hierarchical graph convolutions, separate long- and short-term encoders, and attention-based output. Across three real-world datasets, it consistently outperforms baseline and state-of-the-art comparison methods.
Problem
Multi-step passenger demand forecasting is challenging because future demand depends on nonlinear, dynamic relationships across time and regions.
Method
STG2Seq uses a graph-based sequence-to-sequence model with hierarchical GCN modules, long- and short-term encoders, and attention-based output.
Results
Across three real-world datasets, STG2Seq consistently outperforms baseline and state-of-the-art comparison methods by a large margin.
Takeaways & Limitations
The model captures spatial-temporal correlations without relying on RNNs and accounts for dynamic temporal effects through attention.
Takeaways & Limitations
The approach addresses spatial correlation under region partitions and uses only the most recent h demand steps as historical input.
Abstract
from arXiv · showhide
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.
1 Introduction
On-demand vehicle sharing requires accurate citywide multi-step demand forecasts, but nonlinear spatial-temporal dependencies make prediction difficult. STG2Seq addresses these issues with a graph-based sequence-to-sequence architecture and reports consistent gains over comparison methods.
- Low demand near drivers can cause long searches, while high demand near passengers can cause ride delays, creating waiting-time, income, and energy costs.
- Passenger demand forecasting must capture nonlinear, dynamic dependencies across both a target region’s history and demand in other city regions.
- CNN-based methods impose grid partitions and cannot represent non-Euclidean correlations between remote regions with similar characteristics.
- RNN-based architectures can lose long-term temporal information through iterative processing and accumulate errors during multi-step decoding.
- STG2Seq formulates citywide demand as a graph and uses hierarchical GCN modules, long- and short-term encoders, and attention-based output modeling.
- Experiments on three real-world datasets show that STG2Seq consistently outperforms three baselines and eight deep-learning state-of-the-art methods by a significant margin.
2 Notations and Problem Statement
The task is to forecast citywide passenger demand for the next τ time steps from recent demand history and time features. Demand is represented across N city regions at each time step.
- At each time step, a matrix D_t represents passenger demand across N regions, while vector E_t contains time-of-day, weekday, and holiday features.
- The prediction function Γ forecasts the next τ citywide demand steps using the most recent h demand steps and available time features.
3 Methodology
STG2Seq uses graph-based encoders and gated graph convolution to jointly model spatial-temporal demand patterns, then combines long- and short-term representations with temporal and channel attention for multi-step forecasting.
- The architecture contains long-term and short-term encoders plus an attention-based output module, with GGCMs in both encoders.GGCMs extract spatial-temporal correlations simultaneously through GCN along the temporal axis.
- 3.1 Passenger Demand on Graph: The city is represented as a graph whose nodes are regions, with connectivity defined by similarity between passenger-demand patterns.The adjacency matrix is constructed from regional demand-pattern similarity rather than geographic proximity alone.
- 3.2 Long-term and Short-term Encoders: The long-term encoder processes the most recent h citywide demand steps, while the short-term encoder uses a sliding window of q recent steps for each target horizon.The short-term encoder integrates already predicted demand during multi-step forecasting, with the last-step prediction fed back into it.
- 3.3 Gated Graph Convolutional Module: Stacked GGCMs form a hierarchical structure that captures spatial-temporal correlations across the full input window.Each GCN operates on a k-step historical window, and serial GGCMs extend coverage over h or q steps.
- 3.3 Gated Graph Convolutional Module: Within a GGCM, a gating mechanism models nonlinear demand relationships, while residual connections help avoid network degradation.The gate controls which parts of the linear transformation pass through to the output.
- 3.4 Attention-based Output Module: The output module concatenates long-term and recent representations, then applies temporal and channel attention before decoding predictions.Temporal attention assigns time-varying importance to historical steps using target-time features; channel attention identifies important representation channels.
4 Experiments
Experiments evaluate STG2Seq on three real-world demand datasets against traditional and state-of-the-art baselines. The model performs strongly in next-step and multi-step forecasting, while component and irregular-region analyses examine its design and flexibility.
- Datasets: The experiments use DidiSY, BikeNYC, and TaxiBJ, with hourly or 30-minute demand intervals and held-out final days for testing.DidiSY and BikeNYC use hourly steps, while TaxiBJ uses 30-minute steps.
- Next-step prediction: STG2Seq consistently achieves the best next-step performance across all three datasets against traditional and state-of-the-art methods.Relative RMSE improvements are 2.6%, 4.9%, and 2.3% across the three datasets; the cited passage also reports MAE and additional improvements.
- Multi-step prediction: For three-step forecasting, STG2Seq maintains good prediction across all horizons and deteriorates more slowly than ConvLSTM and DCRNN.HA has large but consistent error, whereas ConvLSTM and DCRNN deteriorate rapidly, especially on DidiSY.
- Component analysis: Removing the short-term encoder, attention modules, gating mechanism, or teacher forcing produces variants used to assess the model’s component contributions.The experiments compare five such variants on the BikeNYC dataset.
- Component analysis: Temporal and channel attention improve prediction accuracy and slow multi-step deterioration, while gating models nonlinearities better than ReLU.These observations support the reported design of the attention modules and GGCM gate.
- Irregular regions: STG2Seq can be applied without modification when Shenyang is repartitioned into irregular regions based on the road network.The irregular-region experiment uses precise GPS locations from DidiSY service requests.
5 Conclusion
The conclusion presents STG2Seq as a graph-based framework for multi-step citywide passenger demand forecasting. It combines hierarchical graph convolution, long- and short-term encoders, and attention, and reportedly outperforms state-of-the-art methods on three real-world datasets.
- 5 Conclusion: STG2Seq formulates citywide passenger demand on a graph and uses hierarchical graph convolution to extract spatial and temporal correlations simultaneously.The framework targets multi-step citywide passenger demand forecasting.
- 5 Conclusion: The long-term and short-term encoders enable multi-step prediction without relying on recurrent neural networks.The conclusion identifies these encoders as part of the proposed framework.
- 5 Conclusion: An attention mechanism models dynamic attributes in temporal correlations.The conclusion connects attention with the dynamic aspect of temporal relationships.
- 5 Conclusion: Experiments on three real-world datasets show that STG2Seq outperforms other state-of-the-art methods by a large margin.