The field of spatio-temporal prediction is rapidly advancing, with a focus on developing innovative models for environmental and urban applications. Recent research has emphasized the importance of accurately predicting air quality, traffic emissions, and human mobility patterns.
One of the key trends in this area is the integration of deep learning techniques, such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, to capture complex spatio-temporal dependencies. These models have shown significant improvements in prediction accuracy and generalization capabilities.
Another notable direction is the development of scale-disentangled spatio-temporal modeling frameworks, which aim to decompose and fuse features at different scales to improve long-term prediction performance. These frameworks have demonstrated state-of-the-art performance in traffic emission forecasting and other applications.
Noteworthy papers in this area include:
- STPFormer, which proposes a pattern-aware spatio-temporal transformer for traffic forecasting, achieving state-of-the-art performance on multiple real-world datasets.
- STM3, which introduces a mixture of multiscale Mamba architecture for long-term spatio-temporal time-series prediction, demonstrating superior performance on real-world benchmarks.
- GeoMAE, which presents a novel contrastive self-learning framework for spatio-temporal graph forecasting with missing values, outperforming models trained from scratch on real-world datasets.