NEST framework tackles dataset shifts with regime-oriented MoE
New approach beats state-of-the-art on network traffic and physics data.
NEST, introduced by Lanhao Li, Bingshu Xie, Lijun Sun, Xin Xue, Haoyi Zhou, and Jianxin Li, addresses a critical gap in long-term forecasting: dataset-level distribution shifts caused by multiple underlying operational regimes. Unlike prior work that focuses only on local temporal shifts, NEST explicitly models the global structural challenge by first partitioning the dataset into distinct regimes through unsupervised clustering in a moment-entropy space. It then uses a regime-oriented router mechanism that generates initial expert weights from temporal content, refined via geometric modulation toward regime centroids. Individual experts act as specialized kernels, learning unique variate-attention patterns for each regime.
Evaluated on diverse benchmarks including heterogeneous network traffic and physical phenomena, NEST consistently achieves state-of-the-art performance. The code and datasets are publicly available, enabling further research and application in any domain where time-series data evolves across different behavioral modes—such as climate modeling, energy load forecasting, and financial market analysis. This work represents a significant step toward more robust and interpretable forecasting models.
- NEST uses a two-phase dense Mixture-of-Experts architecture to explicitly model dataset-level distribution shifts.
- It partitions data into operational regimes via unsupervised clustering in a moment-entropy space.
- Achieves state-of-the-art results on heterogeneous benchmarks including network traffic and physical phenomena.
Why It Matters
Enables robust long-term forecasting in complex systems by explicitly handling regime changes, improving reliability in climate, energy, and finance.