Researchers have introduced ScaleIn, a novel training methodology for time series foundation models (TSFMs) designed to address issues arising from significant scale variations across datasets. Existing methods like Reversible Instance Normalization (ReVIN) can inadvertently weight series by their scale, leading to ScaleCon, where larger-scale series dominate training. ScaleIn, however, ensures that the optimization trajectory is invariant to arbitrary rescaling of training series, regardless of the loss function used, provided it's homogeneous of degree p. This approach has demonstrated substantial improvements, reducing Mean Absolute Scaled Error (MASE) by an average of 18.8% on GIFT-Eval and 21.9% on M-competitions across various TSFM architectures and benchmarks. AI
IMPACT This new training method, ScaleIn, significantly improves the performance of time series foundation models, potentially leading to more accurate forecasting and analysis across various domains.
RANK_REASON The cluster contains an academic paper detailing a new methodology for training time series foundation models. [lever_c_demoted from research: ic=1 ai=1.0]
- GIFT-Eval
- Hugging Face
- MAE
- M-competitions
- Microsoft Security Essentials
- Quantile Loss
- Reversible Instance Normalization
- ScaleCon
- ScaleIn
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
