Fuzzy-MoE Enhances Time Series Forecasting with Interpretable Expert Routing.

Forecasting non-stationary multivariate time series is challenging because different variables and time segments often exhibit unique underlying dynamic states. Traditional deep learning models typically compress these diverse dynamics into a single, end-to-end mapping, which can lead to suboptimal predictions and a lack of transparency regarding how the model arrives at its forecasts. To address these limitations, researchers propose Fuzzy-MoE, a novel dynamic Mixture-of-Experts (MoE) model grounded in fuzzy logic. Fuzzy-MoE reframes time series forecasting as a process of identifying latent temporal states and then interpretably routing data to specialized expert networks. The core of the system is a dual-view fuzzy router that analyzes both local convolutional dynamics and global segmented statistics. This router employs learnable Gaussian membership functions to infer latent temporal states and calculate the activation strength for each expert. This enables explicit, IF-THEN rule-based expert selection, allowing different variables within the same time series to be processed by different experts. This fine-grained, interpretable routing strategy not only captures heterogeneous temporal dynamics more effectively but also significantly improves forecasting accuracy compared to mainstream methods, while providing clear diagnostics for expert selection.