Causal Conditional Hidden Markov Model for Multimodal Traffic Prediction

Multimodal traffic flow can reflect the health of transportation system, and its prediction is crucial to urban management. Recent works overemphasize spatio-temporal correlations flow, ignoring physical concepts that lead generation observations their causal relationship. Spatio-temporal are considered unstable under influence different conditions, spurious may exist in observations. In this paper, we analyze affecting multimode from perspective observation principle propose a Causal Conditional Hidden Markov Model (CCHMM) predict multimodal flow. latent variables inference stage, posterior network disentangles representations interest conditional information observations, propagation module mines data prior samples distribution feeds them into generator generate We use mutually supervised training method for enhance identifiability model. Experiments on real-world datasets show CCHMM effectively disentangle identify causality, accurately