Reproduce “End-to-End Neural Decomposition with Koopman Operators for Time-Series Forecasting” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
De-Yan Lu, Xugang Lu, Yu Tsao, Jian-Jiun Ding
A neural framework called NDKoop jointly models dynamics with Koopman operators and performs learnable signal decomposition for time-series forecasting. It decomposes signals into trend and periodic components, each governed by appropriate Koopman operators, achieving superior prediction accuracy and computational efficiency across various real-world datasets.
Key insights
- Jointly learning signal decomposition and Koopman dynamics end-to-end enables a data-driven, adaptive approach to modeling complex time series.
- Modeling different signal components (trend vs. periodic) with distinct types of Koopman operators (frequency-independent vs. frequency-dependent) effectively captures nonstationary and frequency-varying behaviors.
- Approximating traditional signal decomposition methods (like VMD) with neural networks (NVMD) integrates them efficiently into an end-to-end deep learning framework, balancing accuracy and computational cost.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.