Interpretable time-frequency collaboration and multi-scale learning for adaptive classification of long time series

The coupling of multi-scale temporal-frequency structures in long time-series data increases the difficulty of representation learning, while inconsistency in cross-domain distributions further limits the stability and generalization ability of traditional single-scale or single-view models. To address the practical demand for cross-domain transfer and robust classification of long time-series data in financial scenarios, this paper proposes an adaptive multi-scale time-frequency collaboration model, namely MS-TFCM (Multi-Scale Time-Frequency Collaboration Model). The model first employs a lightweight multi-scale temporal encoder based on three causal 1D convolutional branches, which capture local shocks, medium-term oscillations, and long-term trends with linear complexity in sequence length. Meanwhile, a tri-band frequency decomposition module is introduced to explicitly characterize low-frequency trends, mid-frequency oscillations, and high-frequency disturbance components. Subsequently, the temporal and frequency representations are mapped into a shared latent space, where time-frequency collaborative constraints and maximum mean discrepancy alignment are jointly imposed through gated fusion to enhance cross-domain adaptation under distribution shift. Finally, a post-hoc explanation mechanism based on structural decomposition is constructed to quantitatively analyze the trend contribution, band contribution, and residual contribution in the prediction results. Experimental results on a cross-market financial dataset demonstrate that the proposed model effectively improves cross-domain classification performance by up to 3.7% in accuracy, and provides decomposition-level interpretive evidence for prediction results.