DST-Net decomposes trend and seasonal time series components

DST-Net uses a disentangled architecture to apply different inductive biases to time series components: a Latent Graph Block for spatial-heavy trend modeling and a Gaussian Multi-scale Attention mechanism for temporal-heavy seasonal components. This approach addresses the distinct structural properties of trend versus seasonal data.