Consistency regularization for few shot multivariate time series forecasting.

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Multivariate time series forecasting requires extensive, high-quality data for model generalizability, which is often difficult to obtain in practice.
• Existing methods struggle with data scarcity and capturing complex inter-variable dependencies across different frequencies.

Purpose of the Study:

• To develop an advanced algorithm for multivariate time series forecasting that addresses data limitations and improves predictive performance.
• To enhance model generalizability and accuracy by effectively utilizing limited training data.

Main Methods:

• The proposed algorithm combines time-frequency mining with consistency regularization to augment training data via weak perturbation techniques.
• Employs consistency regularization to ensure model stability against input data variations, creating richer training samples.
• Utilizes two complementary dependency extractors to adaptively capture variable interactions across different frequency patterns.

Main Results:

• The method demonstrates improved ability to learn diverse data patterns and features through enhanced training samples.
• Adaptive capture of frequency-specific interactions enhances the model's perception and processing of time series data.
• Validation on five real-world datasets shows superior performance compared to existing multivariate time series forecasting methods.

Conclusions:

• The proposed algorithm effectively overcomes data scarcity challenges in multivariate time series forecasting.
• Combining time-frequency mining and consistency regularization leads to more robust and accurate forecasting models.
• The method offers a significant advancement in predicting future trends from complex, multivariate time series data.