Meta-Learning Enhanced Trade Forecasting: A Neural Framework Leveraging Efficient Multicommodity STL Decomposition

Figure 2

Architecture of the Dual-channel Spatiotemporal Encoder showcasing the integration of temporal attention mechanism, dilated causal convolution, and Struc2Vec graph embedding to capture both the temporal correlations and spatial relationships in trade data. The encoder effectively processes and integrates trend and seasonal components decomposed via STL, facilitating accurate trade forecasting by leveraging global and local patterns.