Multi-Modal Time Series Prediction via Mixture of Modulated Experts

In response to the limitations of existing multimodal time series prediction methods in the face of scarce high-quality paired data and difficult feature alignment, researchers have proposed a new mechanism called Expert Modulation. This mechanism combines modulation experts, conditioning both routing and expert computation on text signals, thereby achieving direct and efficient control over cross-modal expert behavior. Theoretical analysis and experimental results show that this method performs significantly better than existing methods in multimodal time series prediction tasks. The relevant code has been open-sourced on the GitHub platform.