Enhanced Dynamic Hypergraph Neural Networks for Time Series Prediction

Accurately capturing high-order correlations and temporal dependencies in multivariate time series is a crucial challenge in data-driven modeling, particularly when integrating heterogeneous data sources. To address these challenges, this study proposes a novel Dynamic Hypergraph Forecasting Network (DHFN). The framework dynamically constructs hypergraphs based on Pearson correlation coefficients between features, where hyperedges are adaptively aggregated using a hypergraph convolutional network to extract high-order feature relationships. These dynamically learned feature representations are further processed by a GRU network to capture temporal dependencies, providing a unified framework for multivariate time series forecasting. The proposed DHFN is evaluated on diverse datasets, demonstrating superior performance in terms of both MAE and RMSE compared to baseline methods such as CNN-GRU and traditional hypergraph approaches. The results highlight the potential of the DHFN framework as a generalizable solution for dynamic feature integration and multivariate time series modeling.