FREQMIXATTNET: CONTRASTIVELY SUPERVISED FREQUENCY-MIXING ATTENTION FOR TIME-SERIES FORECASTING

Wei li, Zhumei Gou, Xin Cai, Wanjie Tao, Yanqiu Zhang

Keywords: Time series, Frequency-domain, Attention, contrastive learning

TL;DR: FreqMixAttNet: Cross-domain time-frequency fusion framework using adaptive convolution and contrastive auxiliary loss for superior time series forecast

Abstract: Time series forecasting has gained significant attention due to its wide applicability in domains such as traffic prediction and weather monitoring. However, it remains a challenging task because of complex temporal patterns, such as multi-scale periodicities and dynamic fluctuations. Existing methods often focus on either time-domain decomposition or frequency-domain analysis, but rarely integrate both effectively.In this paper, we propose FreqMixAttNet, a novel cross-domain forecasting framework that mixes time and frequency representations via a cross-domain attention mechanism. We first introduce an adaptive convolutional wavelet decomposition to model and separate trend and seasonal components more efficiently. The seasonal part is dual-encoded in both time and frequency domains, which are treated as distinct modalities and fused through a cross-transform attention module. Meanwhile, the trend component is captured by a simple multi-scale MLP in the time domain.To further enhance robustness without pretraining, we incorporate a contrastive auxiliary loss. The combination of adaptive convolution, cross-domain mixing attention, and contrastive learning contributes to the superior performance of our method. Extensive experiments on multiple real-world benchmarks show that FreqMixAttNet consistently outperforms prior state-of-the-art methods, demonstrating the effectiveness of our unified cross-domain design.

Primary Area: learning on time series and dynamical systems

Submission Number: 8746