Driven by the digital economy, personalized recommendation methods are studied to address the problem of information overload in response to changes in user behavior patterns and preference analysis needs on cross-border e-commerce platforms, which can help improve user shopping efficiency and satisfaction. This article proposes a Deep Incremental Recommendation (DIR) model, which is co trained with the basic model and incremental model. By utilizing the prediction error of consumer behavior preference historical data and the sorting error of incremental data on e-commerce platforms to optimize parameters, combined with Dynamic Convolution, the HGNetV2 backbone network is improved to enhance feature extraction ability and model generalization. The experiment was based on a publicly available dataset on Taobao, and compared with benchmark models such as BPR, CF, and CFN. The experimental results showed that the proposed method improved accuracy by 35.5% (0.2387-0.2439) when the number of users increased from 10 to 60, with a stable recall rate of over 15% and an F1 score of 0.1854 (61.5% higher than CFN); When the number of recommended products K=10, the accuracy, recall, and F1 value all reach their peak values, and the curve is smooth with high convergence efficiency. Research has shown that the proposed model effectively balances precise recommendation and interest coverage through dynamic preference weight allocation and multi-scale feature fusion, significantly improving the personalized recommendation performance of cross-border e-commerce platforms.
