Intelligent Decision-Making Architecture Powered by Decision Transformer for Cross-Border E-Commerce Supply Chains

Abstract

This paper introduces an innovative demand forecasting framework for cross-border e-commerce supply chains, called Trans-Demand Net, specifically designed to address the complexity and uncertainty in demand forecasting for supply chain management. The framework seamlessly integrates the multi-head self-attention mechanism of the Transformer architecture to effectively capture long-term dependencies in time series data, while leveraging GCN to analyze the logistics network structure within the supply chain. Additionally, by incorporating the SAC reinforcement learning strategy, the model's adaptability and decision-making optimization are further enhanced. Experimental results on the Amazon Product Reviews Dataset and Alibaba B2B Transaction Dataset demonstrate that Trans-Demand Net significantly outperforms traditional forecasting models across multiple key performance metrics, including MAE, MSE, and RMSE. Through a series of ablation experiments, we further validate the positive contributions of each component within the framework to the overall model performance. Trans-Demand Net exhibits outstanding predictive accuracy and robustness in the dynamic environment of cross-border e-commerce supply chains, providing an innovative and effective solution for intelligent supply chain management.