NEST: Tackling Dataset-Level Distribution Shifts via Regime-Oriented Mixture-of-Experts

Short summary

NEST is a two-phase dense mixture-of-experts framework that addresses dataset-level distribution shifts in multivariate time-series forecasting by partitioning data into distinct operational regimes via unsupervised clustering in moment-entropy space. A regime-oriented router generates initial expert weights refined through geometric modulation, with individual experts functioning as specialized kernels capturing regime-specific dynamics. NEST achieves state-of-the-art performance on benchmarks including heterogeneous network traffic and physical phenomena, with code and datasets publicly available.

• •NEST uses unsupervised clustering in moment-entropy space to identify distinct operational regimes in time-series data
• •Regime-oriented router with geometric modulation assigns expert weights; experts act as specialized kernels for regime-specific dynamics
• •Outperforms existing methods on network traffic and physical phenomena benchmarks; code is open-source

Generated with AI, which can make mistakes.