Condition-Guided Urban Traffic Co-Prediction With Multiple Sparse Surveillance Data

Traffic prediction is one of the important research directions in Intelligent Transportation Systems, with positive implications for vehicle dispatching and vehicle management. In reality, due to the unreliability of data transmission and the volatility of storage devices, data sparsity limits the stability and prediction performance of existing methods that rely on high-quality observed data. To achieve robust sparse traffic prediction, we first investigate two findings as our motivation: the influence of external geographical features on shaping traffic distribution and the coupling dependence of multi-source urban data. Specifically, we develop a condition-guided collaborative learning network for traffic prediction with sparse data. The core idea is to exploit both the informative external geographic features and multiple urban data as auxiliary information to cooperatively learn mobility patterns from sparse data. First, we design an attention-based bilateral filter, which explicitly models the influence of external geographic features on spatial-temporal targets, and exploits such patterns as conditions to further estimate the missing elements. Secondly, a collaborative-learning framework, including a graph fusion module and a memory-preserved mechanism is devised to adaptively extract and aggregate fragments of similar spatial-temporal sequences from multiple urban data, helping the model to learn comprehensive mobility patterns from sparse data. We verify the excellent effectiveness of our model on multi-source traffic datasets collected in modern urban transportation systems.