Extreme-Aware Time-Series Forecasting via Weak-Label-Guided Mixture of Experts.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Deep time-series forecasting models excel in average accuracy but fail on rare, high-impact extreme events.
• Class imbalance in data biases models toward majority patterns, hindering prediction of critical events like natural disasters or power outages.

Purpose of the Study:

• To develop a novel forecasting approach that reliably predicts rare, high-impact extreme events.
• To enhance model interpretability and audibility for real-world deployment in critical scenarios.

Main Methods:

• Introduced a weak-label-guided mixture of experts (WL-MoE) routing inputs to specialized temporal regime models.
• Implemented a two-stage training curriculum: Stage I uses weak labels for expert specialization, Stage II refines forecasting accuracy.
• Enabled interpretable routing via expert-usage profiling for model behavior auditing.

Main Results:

• WL-MoE reduced average Mean Squared Error (MSE) by 7.9% and extreme-case MSE by 23.58% across seven benchmark datasets.
• In a UK flood forecasting study, WL-MoE decreased all-water MSE by 31.6% and high-water MSE by 35.0%.

Conclusions:

• Weak-label guidance effectively stabilizes expert specialization, improving reliability for rare extreme events.
• The WL-MoE approach enhances forecasting accuracy and provides auditable model behavior for critical applications.