Research map / Time series / Long-term forecasting
BasisFormer: Attention-based Time Series Forecasting with Learnable and Interpretable Basis
Zelin Ni, Hang Yu, Shizhan Liu, Jianguo Li, Weiyao Lin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace learnable basis with fixed sine/cosine basis (ablation) | BasisFormer · Electricity | MSE (avg): +9.3% | Hurt |
| replace learnable basis with random sine/cosine basis (ablation) | BasisFormer · Electricity | MSE (avg): +9.1% | Hurt |
| replace learnable basis with covariate embedding (ablation) | BasisFormer · Electricity | MSE (avg): +5.3% | Hurt |
| add InfoNCE loss to basis module (ablation) | BasisFormer · Electricity | MSE (avg): 0.204 -> -5.2% | Improved |
| add smoothness loss to basis module (ablation) | BasisFormer · Electricity | MSE (avg): 0.204 -> -2.0% | Improved |
| use both InfoNCE and smoothness losses (ablation) | BasisFormer · Electricity | MSE (avg): 0.204 -> -7.2% | Improved |
| increase Coef heads from 4 to 16 (ablation) | BasisFormer · Electricity | MSE (avg): 0.201 -> 0.189 | Improved |
| increase Coef heads from 16 to 32 (ablation) | BasisFormer · Electricity | MSE (avg): 0.189 -> 0.199 | Hurt |
| vary number of bases N from 1 to 10 (ablation) | BasisFormer · Electricity | MSE (avg): 0.196 -> 0.187 | Improved |
| increase number of bases N from 10 to 20 (ablation) | BasisFormer · Electricity | MSE (avg): 0.187 -> 0.190 | No clear effect |
| add self-supervised coef module supervision to FEDformer covariates | FEDformer · Electricity | MSE (avg): 0.214 -> 0.204 | Improved |
| add self-supervised coef module supervision to Autoformer covariates | Autoformer · Electricity | MSE (avg): 0.227 -> 0.214 | Improved |
Closest papers by meaning
- FilterNet: Harnessing Frequency Filters for Time Series Forecasting (NeurIPS 2024 · similarity 0.89)
- TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.89)
- Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts (ICLR 2025 · similarity 0.89)
- Sequence Complementor: Complementing Transformers for Time Series Forecasting with Learnable Sequences (AAAI 2025 · similarity 0.88)
- TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting (NeurIPS 2025 · similarity 0.88)
- TimeXer: Empowering Transformers for Time Series Forecasting with Exogenous Variables (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
