Research map / Time series / Foundation models for time series
CauKer: Classification Time Series Foundation Models Can Be Pretrained on Synthetic Data
Shifeng Xie, Vasilii Feofanov, Jianfeng Zhang, Themis Palpanas, Ievgen Redko
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 TabPFN-style SCM data with CauKer data (ablation) | Mantis/MOMENT pretraining · UCR zero-shot | accuracy: Mantis 73.49 -> 78.31; MOMENT 59.23 -> 74.24 | Improved |
| Add non-zero mean functions to KernelSynth GP (ablation) | KernelSynth · UCR zero-shot | accuracy: Mantis 77.70 -> 78.20; MOMENT 69.31 -> 72.56 | Improved |
| Add SCM causal structure on top of Mean+KernelSynth (ablation) | CauKer · UCR zero-shot | accuracy: Mantis 78.20 -> 78.31; MOMENT 72.56 -> 74.24 | Improved |
| Scale CauKer pretraining data from 10K/100K to 10M (ablation) | MOMENT / Mantis · UCR zero-shot | accuracy: MOMENT 74.24 -> 77.49; Mantis 76.91 -> 79.09 | Improved |
| Scale UEA real pretraining data to full size (ablation) | MOMENT / Mantis · UCR zero-shot | accuracy: Mantis 76.21 (127K) -> 71.93 (12.67M) | No clear effect |
| Increase kernels and parents per node together (ablation) | CauKer · UCR zero-shot | accuracy: 0.7848 -> 0.7810 | No clear effect |
| Use only DotProduct kernels instead of rich kernel bank (ablation) | CauKer · UCR zero-shot | accuracy: 76.79 vs 78.07 (RBF-only) | Hurt |
| Pretrain on 100K CauKer instead of 1.89M real series | Mantis · UCR zero-shot | accuracy: drop less than 0.1% | No clear effect |
| Pretrain on 10M CauKer instead of 13M real series | MOMENT · UCR zero-shot | accuracy: drop barely exceeds 1% | Mixed |
| Add SCM layer to KernelSynth generation | CauKer generation cost · 1000 series, length 512 | generation time: 182.25 -> 121.64 s | Improved |
Closest papers by meaning
- MOMENT: A Family of Open Time-series Foundation Models (ICML 2024 · similarity 0.86)
- Multi-Scale Finetuning for Encoder-based Time Series Foundation Models (NeurIPS 2025 · similarity 0.86)
- Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization (ICML 2025 · similarity 0.86)
- Towards Neural Scaling Laws for Time Series Foundation Models (ICLR 2025 · similarity 0.85)
- Large Pre-trained time series models for cross-domain Time series analysis tasks (NeurIPS 2024 · similarity 0.85)
- Chimera: Effectively Modeling Multivariate Time Series with 2-Dimensional State Space Models (NeurIPS 2024 · similarity 0.85)
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.
