A universal gating framework for multi-expert fusion in heterogeneous multimodal time series forecasting

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Forecasting complex systems traditionally relies on numerical time series data.
• Integrating diverse, heterogeneous data sources (e.g., text, external signals) poses a significant challenge for current multi-modal frameworks.
• Existing methods often use tightly coupled architectures or static fusion, limiting adaptability to evolving data relevance.

Purpose of the Study:

• To develop a novel framework for universal heterogeneous data integration in forecasting.
• To enable seamless fusion of disparate model families (e.g., Large Language Models, numerical forecasters) without architectural modification.
• To introduce adaptive, dynamic weighting of data modalities based on temporal context.

Main Methods:

• Proposed GMM-TS (Gating Mixture-of-Models for Time Series) framework, a modular architecture inspired by Mixture-of-Experts.
• Operates directly in the target prediction space for direct integration of diverse models.
• Employs a Transformer-based gating mechanism for dynamic, per-time-step expert weight computation, replacing static interpolation.

Main Results:

• GMM-TS consistently outperforms state-of-the-art baselines across nine diverse domains.
• Demonstrated superior predictive accuracy and enhanced interpretability by revealing modality importance.
• Successfully integrated more than two expert models, showcasing extensibility.

Conclusions:

• GMM-TS offers an efficient, extensible, and interpretable modality-agnostic solution for multi-source data forecasting.
• The dynamic gating mechanism allows adaptive prioritization of data sources as temporal context evolves.
• This framework advances real-world forecasting capabilities in complex systems with heterogeneous data environments.