DecompKAN: Lightweight Attention-Free Architecture for Time Series Forecasting
A groundbreaking architecture called DecompKAN has been introduced to enhance long-term forecasting for time series data. This innovative model integrates several approaches, including trend-residual decomposition and channel-wise patching, alongside learned instance normalization and B-spline KAN edge functions. Remarkably, DecompKAN has achieved or equaled the top mean squared error metrics across 15 of 32 dataset-horizon combinations when compared to existing benchmarks. It excelled in 20 of 36 evaluations across nine datasets, notably the PPG-DaLiA physiological benchmark. Lightweight in design, DecompKAN omits attention mechanisms, allowing for straightforward visualization of its scalar function. The study is accessible on arXiv under the ID 2604.23968.
Key facts
- DecompKAN combines trend-residual decomposition, channel-wise patching, learned instance normalization, and B-spline KAN edge functions.
- Achieves best or tied-best MSE on 15 of 32 dataset-horizon combinations.
- Achieves best or tied-best MSE on 20 of 36 comparisons under controlled same-recipe evaluation.
- Evaluated on 9 datasets including PPG-DaLiA benchmark.
- Each KAN edge learns an explicit, inspectable 1D scalar function.
- Architecture is attention-free and lightweight.
- Particular strength on datasets with smooth temporal patterns.
- Published on arXiv with ID 2604.23968.
Entities
Institutions
- arXiv