Bike Sharing Demand Prediction based on Knowledge Sharing across Modes: A Graph-based Deep Learning Approach

The paper addresses the limitation that "most existing models of bike-sharing demand prediction are solely based on its own historical demand variation, essentially regarding bike sharing as a closed system and neglecting the interaction between different transport modes. The authors note that bike sharing is often used to complement travel through other modes (e.g., public transit) and that there is no existing method capable of leveraging spatiotemporal information from multiple modes with heterogeneous spatial units," such as station-based subway stations versus stationless ride-hailing zones.

To address this research gap, the study proposes a graph-based deep learning approach for bike sharing demand prediction (B-MRGNN) with multimodal historical data as input. The model architecture "is composed of L multi-relational spatiotemporal blocks (STMR blocks) for multimodal representation learning, each comprising TCNs to model temporal patterns and multirelational graph neural networks (MRGNNs) to model the spatial influence of adjacent subway stations and ride-hailing zones on BSS stations."

The methodology relies on a multi-relational graph construction to encode spatial dependencies within and across modes. Specifically, intra-modal graphs are used to capture spatial correlations among stations/zones of the same mode, while inter-modal graphs are defined to capture the pairwise correlations among stations/zones between the target mode (i.e., bike sharing) and each of the auxiliary modes (i.e., subway and ride-hailing). To capture different types of relationships, the authors define two adjacency matrices for each graph: one for geographical proximity denoted as A G, and the other for semantic similarity denoted as A P.

The MRGNN component is designed to capture correlations between nodes through message passing while addressing the disparity between different modes, which might result in negative transfer. To tackle this, the authors introduce an inter-modal graph convolution network to aggregate information of connected subway stations and ride-hailing zones for each BSS station, considering both cross-mode similarity and difference. The temporal dependencies are captured using a temporal gated convolution network (TCN), where a separate TCN layer for each mode is applied to capture mode-specific temporal information.

To improve training efficiency, the authors introduce a prediction-based regularization term. This involves generating future predictions of subway and ride-hailing demand with additional feed-forward networks, and use the prediction error of subway and ride-hailing demand as a regularization term in the loss function. The resulting loss function is:

L(theta) = bt+1 - X bt+1 + epsilon r sum m in s,h mt+1 - X mt+1

The model was validated using real-world bike sharing, subway and ride-hailing data from New York City, specifically using Citi Bike, NYC Subway, and NYC Ride-hailing (FHV) data. Extensive experiments demonstrate that the results demonstrate the superior performance of our proposed approach compared to existing methods, including HA, LR, XGBoost, LSTM, STGCN, MGCN, and Graph WaveNet. Specifically, compared with Grave WaveNet, our proposed model can further reduce the prediction error, with RMSE and MAE improvement of 8.6% and 10.2% respectively. Furthermore, the study found that incorporating either subway or ride-hailing demand patterns can already significantly improve the prediction performance of bike sharing demand, with the RMSE reduced by 6.2% and 7.2% respectively.

1. Dynamic Multi-Relational Graph Learning

• Improvement: Replace the predefined adjacency matrices (A G and A P) with a self-evolving Graph Attention Mechanism that learns topology in real-time.

• What the improved system can do: Instead of relying on static geographical or semantic similarities, the system can detect and adapt to shifting urban patterns, such as temporary road closures, sudden shifts in commuting behavior due to weather, or seasonal changes in how different transport modes interact.

2. Universal Heterogeneous Spatiotemporal Engine

• Improvement: Generalize the B-MRGNN architecture to handle any combination of point-based (stations) and area-based (zones) entities across any number of modalities.

• What the improved system can do: It can be deployed for city-wide logistics orchestration, predicting the interplay between autonomous delivery drones (point-based), ride-hailing fleets (zone-based), and heavy freight trucking (route-based) to optimize urban supply chains and reduce congestion.

3. Probabilistic Multimodal Uncertainty Quantification

• Improvement: Integrate Bayesian Neural Networks or Quantile Regression into the STMR blocks to move from point estimates to probabilistic distributions.

• What the improved system can do: Instead of providing a single demand number, the system provides a confidence interval (e.g., 90% probability that demand will be between 50 and 70 bikes). This allows city planners and operators to perform risk-aware decision-making and prepare for worst-case demand surges.

4. Explainable Cross-Modal Attribution

• Improvement: Incorporate an attention-based interpretability layer within the inter-modal graph convolution network to calculate influence scores between modes.

• What the improved system can do: It can provide actionable insights by explaining why a prediction was made (e.g., The predicted 20% spike in bike-sharing demand is 75% attributed to the subway service disruption at Station X and 25% to the surge in ride-hailing prices in Zone Y).

5. Predictive-to-Prescriptive Reinforcement Learning (RL) Integration

• Improvement: Use the B-MRGNN's multimodal predictions as the state representation for a Deep Reinforcement Learning agent.

• What the improved system can do: It shifts from merely predicting demand to actively managing it. The system can autonomously trigger proactive bike rebalancing, adjust dynamic pricing for ride-hailing, or suggest subway frequency adjustments to mitigate predicted demand imbalances before they occur.