Dynamic TMoE: Drift-Aware Framework for Non-Stationary Time Series Forecasting
A new framework called Dynamic TMoE has been introduced by researchers for forecasting non-stationary time series, specifically targeting distribution shifts. This model employs Maximum Mean Discrepancy (MMD) to identify these shifts, enabling the dynamic creation of diverse experts while eliminating unnecessary ones. Additionally, a temporal memory router utilizes recurrent states alongside an anomaly repository to ensure the selection of stable, context-sensitive experts without requiring updates during testing. Results from experiments across nine benchmarks demonstrate its cutting-edge performance, achieving a 10.4% reduction in mean squared error (MSE).
Key facts
- Dynamic TMoE is a drift-aware mixture-of-experts framework.
- It detects distribution shifts via Maximum Mean Discrepancy (MMD).
- Experts are dynamically instantiated and pruned to optimize capacity.
- A temporal memory router uses recurrent states and an anomaly repository.
- No test-time updates are required for expert selection.
- Experiments on nine benchmarks show state-of-the-art performance.
- MSE is reduced by 10.4%.
- The framework unifies architectural evolution with temporal continuity.
Entities
—