Decomposition-Based Time-Frequency Augmentation for Few-Shot Time-Series Forecasting

As a pivotal technique for predicting future temporal variations through historical data analysis， time-series forecasting is critical in scientific and engineering domains such as energy management， traffic-flow prediction， financial market-price analysis， and meteorological simulation. Whereas the integration of deep learning has considerably enhanced the accuracy of predictive models， their high reliance on large-scale annotated data continues to impose significant constraints on real-world applications， particularly under few-shot scenarios. Although few-shot learning for vision and language modalities has progressed significantly， the methodological framework for few-shot time-series forecasting remains underdeveloped. The unique temporal-domain characteristics （trend patterns） and frequency-domain features （seasonality） inherent in time-series data render it challenging to directly transfer few-shot learning strategies from other modalities. Hence， we propose a temporal-frequency domain data-augmentation framework for few-shot time-series forecasting， which enhances data diversity through decoupling and reinforcing trend-seasonal components. First， we employ time-series decomposition to decouple raw sequences into trend and seasonal components. Second， we design a temporal mix-up strategy to apply linear interpolation perturbations on trend components， coupled with a dominant-shuffle method that injects controllable noise into the spectral domain of seasonal components. Finally， we construct three augmented samples through component recombination， i.e.， time-augmented， frequency-augmented， and hybrid time-frequency-augmented samples. Extensive experiments on four benchmark datasets demonstrate that our method significantly improves few-shot performance across state-of-the-art forecasting models. Compared with large-language-model-based baselines， our approach enables lightweight models to achieve superior prediction accuracy in 80% of scenarios while substantially reducing computational costs.