Dynamic Tangent Search-Driven Graph Neural Networks for Cross- Border E-Commerce Sales Forecasting

Abstract

The rapid expansion of cross-border e-commerce has intensified the need for accurate and computationally efficient sales forecasting models capable of capturing complex user–product interactions and dynamic market behavior. Traditional machine learning (ML) approaches often fail to model multi-hop relational dependencies and volatile browsing patterns across international platforms. To address these challenges, this study proposes a Dynamic Tangent Search–driven Graph Neural Network (DTS-GNN) framework for cross-border e-commerce sales forecasting. The proposed methodology integrates comprehensive data preprocessing, including noise removal, missing value imputation, Min–Max normalization, and one-hot encoding, followed by Principal Component Analysis (PCA) for dimensionality reduction and feature extraction. A Graph Neural Network (GNN) is employed to model relational structures among users and products, while Dynamic Tangent Search (DTS) is used to adaptively optimize graph weights and hyperparameters, improving convergence stability and prediction accuracy. The model is evaluated using a real-world cross-border e-commerce dataset comprising 2,100 records of user browsing behavior, product attributes, and transactional data. Experimental results demonstrate that the proposed DTS-GNN significantly outperforms existing models, achieving a Mean Absolute Error (MAE) of 1.8420, Root Mean Square Error (RMSE) of 2.2165, Normalized RMSE of 0.0568, Mean Absolute Percentage Error (MAPE) of 10.3924, and a correlation coefficient (R) of 0.9448. Additionally, the framework shows improved reliability (24.6), reduced uncertainty (17.3), and faster search time (10.9 s). These results confirm the effectiveness, robustness, and computational efficiency of the DTS-GNN framework for accurate sales forecasting in dynamic cross- border e-commerce environments.