AdaFusionNet: Disentangling and Fusing Asynchronous Patterns for Long-Range Time Series Forecasting

ZhangJian'an, Hongyi Duan, FengRui Liu, Haoyang Liu

Keywords: Long-Range Time Series Forecasting, Deep Learning, Time Series Decomposition, Heterogeneous Architectures, Model Interpretability, Representation Learning

TL;DR: AdaFusionNet learns an interpretable EMA-based adaptive decomposition ($\alpha$) that enables heterogeneous complexity matching (MLP for trend, CNN for residual), with provable optimality and tighter generalization for time-series forecasting.

Abstract:

Long-range time-series forecasting (LTSF) is hard because real sequences superpose asynchronous patterns—slow trends and fast seasonalities. We identify a failure mode of homogeneous architectures, \emph{trend contamination}, where high-frequency dynamics corrupt the learned trend and degrade long-horizon accuracy. We introduce \textbf{AdaFusionNet}, a structured methodology that \emph{Disentangles, Specializes, and Fuses}. First, a learnable projection adaptively separates low- and high-frequency components. Second, heterogeneous streams match model capacity to component complexity (lightweight MLP for trends; patch-wise CNN for residuals). Third, a synergistic fusion block recombines component forecasts and models cross-channel interactions.

On the theory side, we prove: (i) an upper bound on spectral leakage and consistency of the decomposition parameter; (ii) mixed-smoothness approximation gains with additive Rademacher-complexity bounds; (iii) a fusion oracle inequality and Lipschitz stability that yield distributional-robustness certificates (Wasserstein-DRO) and PAC-Bayes uncertainty control; and (iv) split-conformal prediction with distribution-free coverage. Empirically, across eight standard LTSF benchmarks and four horizons (96–720), AdaFusionNet delivers consistently strong—often state-of-the-art—accuracy, with the largest gains at long horizons; ablations validate each stage and the proposed diagnostics for leakage/asynchrony. Code and reproducible scripts will be released upon publication. These results indicate that structuring LTSF as \emph{disentangle $\rightarrow$ specialize $\rightarrow$ fuse} is a robust, scalable alternative to the prevailing homogeneous paradigm.

Primary Area: learning on time series and dynamical systems

Submission Number: 12461