This study proposes a novel dual-layer neural network (DLNN) model designed to address the challenges of data sparsity and dynamic user interest modeling in e-commerce recommender systems. The DLNN model combines a base interaction modeling module with a deep sequence modeling module via a shared knowledge matrix and introduces a time-decaying attentional mechanism to capture the dynamic changes in user preferences. Extensive experiments on the Amazon TV product dataset demonstrate that the DLNN model outperforms existing benchmarks on several evaluation metrics, including AUC, recall, precision, and NDCG. In particular, the DLNN model exhibits an improvement of approximately 1% in AUC compared to the closest benchmark model, BERT4Rec. The ablation experiments validate the effectiveness of the two-layer structure, the time-decayed attention mechanism, and the dynamic adjustment mechanism, with the two-layer structure contributing the most. The model performs particularly well in short list recommendation tasks and is suitable for resource-constrained environments. This study provides new ideas for the development of personalized recommendation technology in the field of e-commerce and provides a reference for future research on user behavior modeling and recommender system design.