Dynamic TMoE: A Drift-Aware Dynamic Mixture of Experts Framework for Non-Stationary Time Series Forecasting
The article presents Dynamic TMoE, a new framework designed for non-stationary time series forecasting. This framework addresses the limitations of existing Mixture-of-Experts models by allowing for dynamic adaptation to distribution shifts. Experiments show that Dynamic TMoE achieves state-of-the-art performance, significantly reducing forecasting errors.
- ▪Dynamic TMoE is a framework that combines architectural evolution with temporal continuity during the learning phase.
- ▪It detects distribution shifts using Maximum Mean Discrepancy (MMD) and dynamically adjusts expert pools.
- ▪The framework has been tested on nine benchmarks, resulting in a 10.4% reduction in MSE and a 7.8% reduction in MAE.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.20678 (cs) [Submitted on 20 May 2026] Title:Dynamic TMoE: A Drift-Aware Dynamic Mixture of Experts Framework for Non-Stationary Time Series Forecasting Authors:Jiawen Zhu, Shuhan Liu, Di Weng, Yingcai Wu View a PDF of the paper titled Dynamic TMoE: A Drift-Aware Dynamic Mixture of Experts Framework for Non-Stationary Time Series Forecasting, by Jiawen Zhu and 3 other authors View PDF HTML (experimental) Abstract:Non-stationary time series forecasting is challenged by evolving distribution shifts that static models struggle to capture.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
