UFGTime: Reforming the Pure Graph Paradigm for Multivariate Time Series Forecasting in the Frequency Domain
Recent advances in multivariate time series forecasting have seen a shift toward a pure graph paradigm, which transforms time series into hypervariate graphs and employs graph neural networks (GNNs) to holistically capture intertwined spatiotemporal dependencies. While promising, this approach faces notable challenges. First, converting time series into hypervariate graphs often neglects essential temporal sequences, which are vital for accurately capturing temporal dependencies. Second, treating the graph as a complete structure can obscure the varying importance of intra- and inter-series connections, potentially overlooking key local patterns. To address these challenges, we introduce a novel hyperspectral graph data structure that embeds sequential order into frequency signals and employs a sparse yet meaningful topological structure. In addition, we propose the \textsc{Ufgtime} framework, featuring a frequency-based global graph framelet message-passing operator tailored to hyperspectral graphs, effectively mitigating the smoothing issue and capturing global insights through sparse connections. Extensive experiments demonstrate that our framework significantly surpasses state-of-the-art methods, excelling in both short- and long-range time series forecasting while achieving superior efficiency. Our code is available at:~\url{https://anonymous.4open.science/r/UFGTIME-E352}.