[Paper Review] Deep trip generation with graph neural networks for bike sharing system expansion

Yuebing Liang, Fangyi Ding+2 more|arXiv (Cornell University)|Mar 20, 2023

Urban Transport and Accessibility4 citations

TL;DR

This paper proposes a graph neural network (GNN)-based approach, Spatial-MGAT, for predicting station-level trip generation in bike-sharing system expansion, leveraging multi-source urban built environment data and localized spatial graphs with attention mechanisms. The model outperforms sequential deep learning baselines by achieving an R² of 0.783 on new stations and 0.848 on existing ones, demonstrating superior performance and interpretability for long-term, structural planning scenarios.

ABSTRACT

Bike sharing is emerging globally as an active, convenient, and sustainable mode of transportation. To plan successful bike-sharing systems (BSSs), many cities start from a small-scale pilot and gradually expand the system to cover more areas. For station-based BSSs, this means planning new stations based on existing ones over time, which requires prediction of the number of trips generated by these new stations across the whole system. Previous studies typically rely on relatively simple regression or machine learning models, which are limited in capturing complex spatial relationships. Despite the growing literature in deep learning methods for travel demand prediction, they are mostly developed for short-term prediction based on time series data, assuming no structural changes to the system. In this study, we focus on the trip generation problem for BSS expansion, and propose a graph neural network (GNN) approach to predicting the station-level demand based on multi-source urban built environment data. Specifically, it constructs multiple localized graphs centered on each target station and uses attention mechanisms to learn the correlation weights between stations. We further illustrate that the proposed approach can be regarded as a generalized spatial regression model, indicating the commonalities between spatial regression and GNNs. The model is evaluated based on realistic experiments using multi-year BSS data from New York City, and the results validate the superior performance of our approach compared to existing methods. We also demonstrate the interpretability of the model for uncovering the effects of built environment features and spatial interactions between stations, which can provide strategic guidance for BSS station location selection and capacity planning.

Motivation & Objective

• To address the challenge of predicting trip generation for new bike-sharing stations during system expansion, where existing models fail due to structural changes and data sparsity.
• To develop a spatial regression framework that captures complex spatial dependencies and interactions between stations without relying on temporal sequences.
• To enhance interpretability by uncovering the influence of built environment features and spatial correlations on demand patterns.
• To provide strategic guidance for station location selection and capacity planning in long-term urban transportation planning.
• To demonstrate the model's effectiveness in real-world BSS expansion scenarios using multi-year NYC data.

Proposed method

• Constructs multiple localized graphs centered on each target station to model spatial relationships in a scalable way.
• Employs an attention mechanism to learn dynamic correlation weights between stations, capturing non-linear spatial interactions.
• Integrates multi-source urban built environment features (e.g., land use, population, POIs) as node features for each station.
• Treats the problem as a generalized spatial regression model, linking GNNs to traditional spatial regression frameworks.
• Trains the model using historical demand data and urban features, with loss optimized via RMSE and R² metrics.
• Uses a simplified evaluation setup with K=6 months and T=6 months to simulate real-world expansion scenarios.

Experimental results

Research questions

• RQ1 Can a GNN-based model outperform sequential deep learning models in predicting trip generation for new bike-sharing stations during system expansion?
• RQ2 How well can the model capture spatial dependencies and interactions between stations without relying on temporal demand sequences?
• RQ3 To what extent can the model’s attention weights reveal interpretable relationships between built environment features and station demand?
• RQ4 How does the model perform on new stations that are not part of the historical network, especially when they are located far from existing ones?
• RQ5 Can the model be generalized to long-term planning scenarios where system structure changes and data sparsity is high?

Key findings

• The proposed Spatial-MGAT model achieves an R² of 0.783 on new station predictions and 0.848 on existing stations, significantly outperforming DDP-Exp (R²=0.325 and 0.621) and MOHER (R²=0.663 and 0.653).
• The model demonstrates superior robustness in data-sparse scenarios, particularly for new stations far from existing ones, where sequential models fail due to lack of nearby historical data.
• Attention weights in the model reveal meaningful spatial correlations, such as influence from nearby stations and POI clusters, enhancing interpretability for planners.
• The model’s performance remains strong even when the system structure changes, unlike sequential models that assume a stable network configuration.
• The approach is interpretable and can guide strategic decisions by identifying key built environment features that drive demand, such as population density and mixed-use development.
• The study confirms that GNNs can be effectively framed as a generalized spatial regression model, bridging traditional spatial statistics and deep learning.