Meta-Learning-Based Spatial-Temporal Adaption for Coldstart Air Pollution Prediction

Air pollution is a significant public concern worldwide, and accurate data-driven air pollution prediction is crucial for developing alerting systems and making urban decisions. As more and more cities establish their monitoring networks, there is a pressing need for coldstart model training with limited data accumulation in new cities. However, traditional spatial-temporal modeling and transfer learning schemes have been challenged under this scenario because of insufficient usage of available source data and suboptimal transferring strategy. To address these issues, we propose a meta-learning-based spatial-temporal adaptation solution for coldstart air pollution prediction. Our approach is a model-agnostic framework that enables a given backbone predictor with adaption ability across different space and time locations. Specifically, it learns a factorization of the available source data distribution and recognizes the target city as one of its components, greatly reducing the data accumulation requirement and providing coldstart capability. Furthermore, we design a novel bidirectional meta-learner that can simultaneously leverage task embeddings learned from data and features constructed based on prior knowledge. We conduct comprehensive experiments on both synthetic and real-world air pollution datasets of four distinct pollutants. The results demonstrate that our proposed method achieves a 5.2% lower 24-hour prediction mean absolute error (MAE) than pretraining and fine-tuning solutions when facing a new city with only 200 hours of data, which empirically verifies the effectiveness of our approach as a coldstart training solution.