Consistency regularization for few shot multivariate time series forecasting

Abstract

Multivariate time series forecasting aims to accurately predict future trends by capturing and analyzing various features of the time series. Adequate training data are crucial for ensuring the model's generalizability. However, obtaining a sufficient amount of high-quality labeled data is often a challenge in practical applications. To address this problem, we propose an algorithm that combines time-frequency mining with consistency regularization for multivariate time-series forecasting. First, we increase the amount of data by employing weak perturbation techniques, expanding the original data space. Additionally, we ensure that the model maintains stable predictions under variations in the input data consistency regularization. This approach provides the model with richer training samples, enabling it to learn and understand more comprehensively the different variations patterns and features in the data. Second, we used two complementary dependency extractors to adaptively capture interactions between variables from different levels of frequency patterns. This improves the model's ability to perceive and process different frequency information. Finally, we validate the generalization and effectiveness of the proposed method on five real-world datasets. Extensive experimental results demonstrate that our method outperforms existing methods in terms of performance.