LMSNet: Variable-aware multi-scale framework for multivariate time series forecasting

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Multivariate time series (MTS) data contains rich hierarchical information across multiple temporal scales.
• Existing multi-scale forecasting methods often use uniform scaling, ignore variable-specific dynamics, and fail to model global context effectively.
• These limitations can lead to inaccurate predictions due to inconsistent temporal pattern modeling and spurious correlations.

Purpose of the Study:

• To introduce LMSNet, a novel Lightweight Multi-Scale graph learning framework for improved multivariate time series forecasting.
• To address the limitations of existing methods by incorporating variable-specific scaling and synergistic global-local information modeling.
• To establish a new state-of-the-art in multi-scale time series forecasting.

Main Methods:

• LMSNet employs a variable-specific scaling mechanism using frequency-domain analysis to tailor analytical windows for each variable.
• It integrates global and local information, concurrently modeling temporal and inter-variable dependencies at each network layer.
• A momentum update mechanism refines deep-layer graph node representations to prevent spurious correlations.

Main Results:

• LMSNet achieves state-of-the-art performance on eight benchmark datasets.
• It demonstrates a 2.94% reduction in Mean Absolute Error (MAE) compared to baseline methods.
• Significantly reduces GPU memory usage by 83% and accelerates training by 60%.

Conclusions:

• LMSNet offers a lightweight and effective approach to multivariate time series forecasting.
• The variable-specific scaling and synergistic dependency modeling significantly enhance prediction accuracy and efficiency.
• LMSNet represents a substantial advancement in handling complex multi-scale time series data.