Multi-Scale Convolution with Optimal Transport Attention Effect on Multivariate Time Series

The paper proposes MSC-OT, a multivariate time-series forecasting architecture that combines inverted embedding, multi-scale convolution, and Sinkhorn optimal transport. Variables are treated as tokens, while convolution is applied to attention score matrices to capture local structures in the induced variate-interaction space. An optimal-transport formulation uses iterative matrix scaling to encourage balanced information flow across variables. A learnable adaptive fusion combines base attention, convolution-enhanced scores, and OT-regularized scores. The abstract reports strong results on ETT, Electricity, Traffic, Solar-Energy, and Exchange-Rate datasets for short- and long-term forecasting, supported by ablation studies.