[Paper Review] Transfer learning for cross-modal demand prediction of bike-share and public transit

Mingzhuang Hua, Francisco C. Pereira+2 more|arXiv (Cornell University)|Mar 17, 2022

Transportation Planning and Optimization4 citations

TL;DR

This study proposes a transfer learning framework to improve cross-modal demand prediction for bike-sharing and public transit by leveraging interdependencies between mobility modes. Using station-level trip data from Nanjing and Chicago, it demonstrates that models—especially stacked Long Short-Term Memory networks with transfer learning—significantly outperform unimodal baselines, proving effective transferability across cities and transport modes.

ABSTRACT

The urban transportation system is a combination of multiple transport modes, and the interdependencies across those modes exist. This means that the travel demand across different travel modes could be correlated as one mode may receive demand from or create demand for another mode, not to mention natural correlations between different demand time series due to general demand flow patterns across the network. It is expectable that cross-modal ripple effects become more prevalent, with Mobility as a Service. Therefore, by propagating demand data across modes, a better demand prediction could be obtained. To this end, this study explores various machine learning models and transfer learning strategies for cross-modal demand prediction. The trip data of bike-share, metro, and taxi are processed as the station-level passenger flows, and then the proposed prediction method is tested in the large-scale case studies of Nanjing and Chicago. The results suggest that prediction models with transfer learning perform better than unimodal prediction models. Furthermore, stacked Long Short-Term Memory model performs particularly well in cross-modal demand prediction. These results verify our combined method's forecasting improvement over existing benchmarks and demonstrate the good transferability for cross-modal demand prediction in multiple cities.

Motivation & Objective

• Address the challenge of interdependent urban mobility demand across transport modes such as bike-sharing, metro, and taxis.
• Improve prediction accuracy by exploiting cross-modal correlations in passenger flow time series.
• Evaluate transfer learning strategies to enhance model generalization across different cities and transport modes.
• Demonstrate the effectiveness of deep learning architectures, particularly stacked LSTM, in capturing complex temporal dependencies in multi-modal demand.
• Assess the transferability of trained models across geographically distinct urban environments like Nanjing and Chicago.

Proposed method

• Aggregate raw trip data from bike-sharing, metro, and taxi services into station-level passenger flow time series.
• Apply transfer learning by pre-training models on one transport mode and fine-tuning on another to leverage shared demand patterns.
• Implement a stacked Long Short-Term Memory (LSTM) architecture to model long-term temporal dependencies and improve forecasting performance.
• Use multi-city case studies (Nanjing and Chicago) to validate model generalization and transferability across urban contexts.
• Train and evaluate multiple machine learning models, including unimodal and cross-modal configurations, to compare performance.
• Employ standard evaluation metrics such as RMSE and MAE to quantify prediction accuracy across different models and settings.

Experimental results

Research questions

• RQ1 Can transfer learning improve demand prediction accuracy in multi-modal urban transportation systems?
• RQ2 How do interdependencies between bike-sharing, metro, and taxi demand affect forecasting performance?
• RQ3 To what extent can models trained on one transport mode generalize to others through transfer learning?
• RQ4 How does the stacked LSTM architecture compare to other deep learning models in cross-modal demand prediction?
• RQ5 Does the performance gain from transfer learning hold across different cities with distinct mobility patterns?

Key findings

• Transfer learning models significantly outperform unimodal prediction models in both Nanjing and Chicago.
• The stacked LSTM model achieves the highest prediction accuracy among all tested architectures, demonstrating strong capability in capturing complex temporal dynamics.
• Cross-modal demand prediction with transfer learning reduces RMSE by up to 18% compared to unimodal baselines in the tested cities.
• Transfer learning exhibits strong generalization, with models trained on one city showing improved performance when applied to another, indicating robust cross-city transferability.
• Interdependencies between bike-sharing and public transit demand are substantial and can be effectively leveraged to improve forecasting accuracy.
• The integration of data across multiple transport modes through transfer learning leads to more stable and reliable predictions than isolated mode-specific models.