Reproduce “MFF-FTNet: Multi-scale Feature Fusion across Frequency and Temporal Domains for Time Series Forecasting” on OpenResearch

About this paper

Yangyang Shi, Qianqian Ren, Yong Liu, Jianguo Sun

MFF-FTNet presents a framework for time series forecasting that integrates multi-scale features across frequency and temporal domains using contrastive learning and adaptive noise augmentation. The model reduced average Mean Squared Error (MSE) by 7.7% and Mean Absolute Error (MAE) by 4.0% in multivariate forecasting, demonstrating robustness against noise and missing data on real-world datasets.

Key insights

• Fusing multi-scale features from both frequency and temporal domains provides a richer and more comprehensive representation of time series data, leading to improved forecasting accuracy.
• Adaptive noise augmentation and domain-specific contrastive learning mechanisms enhance a model's stability and generalization capabilities in the presence of noisy or incomplete data.
• The use of multi-scale convolutional kernels within the temporal domain, especially larger kernel sizes, is crucial for effectively capturing long-term dependencies and global trends.

Summary by alphaXiv.