Abstract
Multimodal traffic prediction is essential for urban planning and traffic management. However, existing models often rely on unstable spatio-temporal correlations that fail under changing environmental conditions. This paper introduces the Causal Conditional Hidden Markov Model (CCHMM), a framework that disentangles traffic data into high-level physical concepts and models their causal relationships. By incorporating a Structural Causal Model (SCM) into the latent layers and employing a dual-network architecture for mutual supervision, CCHMM achieves state-of-the-art accuracy and robustness. Evaluations on real-world datasets from Beijing demonstrate significant improvements, particularly in speed prediction, where the model outperforms existing baselines by approximately 10%.
Introduction
Traditional traffic prediction models are primarily "correlation-hungry," assuming that historical patterns between different transport modes (e.g., taxis, bikes, buses) will persist. Such models are vulnerable to "spurious correlations" triggered by environmental shifts. For instance, while taxi and bike usage might typically correlate during rush hours, a sudden rainstorm causes bike demand to drop while taxi demand surges.
To address this, we propose moving beyond simple pattern matching toward understanding the underlying causal mechanisms. We identify three core factors governing traffic dynamics: regional attraction (why people travel to a specific area), demand factors (how they choose a mode of transport), and speed factors (how volume impacts flow).
Methodology
The Causal Conditional Hidden Markov Model (CCHMM) integrates causality into the latent state representation. Unlike standard HMMs, CCHMM disentangles its internal variables to represent specific semantic concepts: attraction, bike demand, taxi demand, bus demand, and speed.
Structural Causal Model (SCM)
The model utilizes an SCM to define the interactions between these disentangled variables. For example, it explicitly models the causal link where vehicle volume (cause) influences traffic speed (effect).
Dual-Network Architecture
The framework consists of two primary components:
- PosteriorNet: Infers the latent state by considering external conditions (weather, time, POI) alongside actual traffic observations.
- PriorNet: Predicts the latent state using only external conditions and historical context.
These networks are trained via mutually supervised learning, forcing the PriorNet to align with the PosteriorNet. This process "anchors" latent variables to real-world physical concepts, ensuring the model learns physical rules rather than data noise. Spatial and temporal dependencies are further captured using a Graph Gated Recurrent Unit (GraphGRU).
Experiments and Results
The model was evaluated using two datasets from Beijing's Xicheng District (XC-Trans and XC-Speed), covering June to December 2021. These datasets include inflow/outflow for three transport modes and average road speeds, augmented with meteorological and POI data.
Performance Comparison
CCHMM was compared against several state-of-the-art baselines, including HMM, CCRNN, DMSTGCN, AGCRN, and DGCRN. As shown in the table below, CCHMM consistently outperformed all baselines across all metrics (MAE, RMSE, MAPE) for every transport mode.
Notably, the model achieved a significant 10% improvement in speed prediction. This suggests that the causal relationship between vehicle volume and road speed is highly identifiable and beneficial for predictive accuracy. Ablation studies confirmed that both the causal propagation module and conditional information are critical to the model's success.
Conclusion
The Causal Conditional Hidden Markov Model represents a shift from correlation-based to causation-based traffic forecasting. By encoding physical rules and disentangling latent variables, the model remains robust against environmental anomalies like sudden weather changes. Future work will focus on "causal discovery" to automatically learn causal structures without pre-defined domain knowledge and optimizing the model's computational complexity for large-scale real-time deployment.
Summary updated