Introduction
Predicting bike-sharing demand is a critical task for urban management, ensuring that stations are neither "congested" nor "hungry." While deep learning models like Graph Convolutional Networks (GCNs) have improved spatio-temporal forecasting, they often struggle with spurious correlations. For instance, a model might incorrectly learn that "Mondays always have low usage" if the training data primarily contains rainy Mondays. This reliance on correlation rather than causation makes models fragile during "special scenarios" like extreme weather or holidays.
As shown in Figure 1, external contextual conditions ($C_t$) act as confounders that bias the relationship between input flow ($X_t$) and the hidden states ($H_t$). This paper introduces STNSCM, a framework that leverages causal inference—specifically the frontdoor criterion—to disentangle these environmental factors from intrinsic traffic patterns.
Methodology
The authors propose the Spatio-Temporal Neural Structural Causal Model (STNSCM), which formalizes the prediction task as a Structural Causal Model (SCM).
Frontdoor Criterion Intervention
The core innovation is the application of the frontdoor criterion. Since context ($C_t$) affects both the input and the hidden state, the model isolates the true causal effect of traffic flow by intervening on the causal path. This de-biasing mechanism allows the model to extract features that are robust to external fluctuations.
Model Architecture
The STNSCM architecture, detailed in Figure 2, consists of several key components:
- Dynamic Causality Generator: Instead of a static city map, the model learns a time-varying causal graph ($A_t^{dyn}$) using a squeeze-excitation mechanism to capture evolving inter-node influences.
- Spatio-temporal Neural Structural Causal Unit: This unit processes the de-biased inputs through gated mechanisms to update the hidden states.
- Counterfactual Representation Reasoning: To improve multi-step forecasting, the model performs "what-if" analysis, extrapolating factual states into future counterfactual scenarios (e.g., predicting flow changes if weather conditions shift).
Experimental Evaluation
The model was evaluated on two large-scale datasets: NYC-Bike (New York City) and BJ-Bike (Beijing). Both datasets include 30-minute interval flow data alongside granular weather and temporal information.
Key Results
STNSCM consistently outperformed state-of-the-art baselines, including STGCN and DGCRN:
- NYC-Bike: Achieved an MAE of 2.670, compared to the best baseline's 2.739.
- BJ-Bike: Achieved an MAE of 12.518, compared to the best baseline's 12.9.
Beyond average performance, the model demonstrated exceptional robustness in the "long tail" of the data. During a sudden snowstorm, standard models predicted typical rush-hour flows, while STNSCM correctly identified the causal impact of the weather, leading to accurate low-flow predictions.
Conclusion
By integrating causal inference into spatio-temporal neural networks, STNSCM moves beyond simple pattern matching. Using causality as a regularizer prevents the network from memorizing coincidental patterns, resulting in a more reliable system for intelligent transportation management. Future work may explore the generalization of this approach to other domains like ride-hailing or traffic speed prediction.
Summary updated