Test-Time Training for Spatial-Temporal Forecasting

Publisher:
SIAM
Publication Type:
Conference Proceeding
Citation:
Proceedings of the 2024 SIAM International Conference on Data Mining, SDM 2024, 2024, pp. 463-471
Issue Date:
2024-01-01
Full metadata record
Despite the recent success of deep neural networks in spatial-temporal forecasting, existing methods suffer from distribution shifts between the training and test data, failing to address the non-stationary and abrupt changes at test time. To solve this problem, we propose a novel test-time training framework for spatial-temporal forecasting. Instead of employing a fixed trained model, we adapt the trained model with only one or a mini-batch of test examples to address the test data shifts. The unique spatial structure with hundreds of geographical locations offers an effective batch size to explore the test-time distribution and avoid overfitting. To implement test-time training on spatial-temporal data, we devise a bidirectional cycle-consistent architecture consisting of a forward and a backward cyclic network. Each network has a shared encoder and two direction-aware decoders. At the test time, two self-supervised auxiliary tasks (forward→backward and backward→forward reconstruction) are proposed to adapt the trained model without accessing the target labels. Besides, the bi-cyclic structure of our model can also improve the forecasting task at training time, and ensure consistency between the training and test time. Comprehensive experiments are performed on various spatial-temporal forecasting datasets, demonstrating the effectiveness of the test-time training framework and the bidirectional-cyclic structure.
Please use this identifier to cite or link to this item: