Reproduce “Feature-aligned N-BEATS with Sinkhorn divergence” 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
Joonhun Lee, Myeongho Jeon, Myungjoo Kang, Kyunghyun Park
Feature-aligned N-BEATS uses N-BEATS’s stacked forecast-and-residual architecture for forecasting unseen domains without target-domain examples. It aligns source-domain representations at each stack using Sinkhorn divergence while training on source forecasts. The authors report competitive cross-domain performance, but alignment does not always improve the underlying architecture: FRED out-of-domain N-BEATS-G has sMAPE 0.172 with alignment versus 0.150 without it.
Key insights
- Stack-wise feature alignment effectively enforces domain invariance within hierarchical time series architectures, offering a balance between semantic coverage and mitigating potential gradient issues.
- The Sinkhorn divergence provides a computationally efficient and stable method for aligning complex feature distributions in deep neural networks, demonstrating comparable performance to the more intensive Wasserstein-2 distance.
- Normalizing feature representations prior to alignment is crucial for learning invariant features, enabling the model to focus on consistent patterns like seasonality and trend across domains rather than scale.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.