Ddiom: Differentiated Dependency Modeling for Time Series Forecasting via Multiscale Individualized Channel-wise Operation and Global Feature Mapping

Le résumé fourni par la source

Multivariate time series forecasting struggles with modeling interdependencies across multiple channels, particularly with dynamic and heterogeneous data. Traditional models fail to capture channel-specific characteristics and long-range dependencies, limiting their effectiveness in complex tasks. To address these issues, we propose Ddiom, a novel framework with two key innovations: frequency-domain kernel extraction and scale-aware prediction fusion. The former improves global pattern recognition through frequency analysis, while the latter enhances prediction accuracy by balancing channel dependencies across scales. Experimental results show that Ddiom outperforms state-of-the-art methods in accuracy, robustness, and computational efficiency, making it a promising tool for real-world forecasting in autonomous systems and intelligent decision-making.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.