Accurate personalization in e-commerce is challenged by high-dimensional, time-varying customer data and rapidly shifting behavioral patterns. We proposed a dynamics-aware neural modeling framework that integrated a gated recurrent unit (GRU) encoder with a temporal attention mechanism (TAM) to capture learning dynamics, stability, and generalization across business stages. Sparse, multi-source interaction streams were embedded to reduce feature dimensionality before sequence modeling. The GRU extracted long- and short-term dependencies, while TAM assigned time-step-specific weights to highlight behaviorally salient periods for prediction. To handle distributional drift, model parameters were updated dynamically, ensuring alignment with evolving customer behavior. We evaluated convergence during training, temporal prediction stability, and cross-period generalization. For customer-demand forecasting, the method achieved an accuracy of 0.924, a recall of 0.910, and an F1-score of 0.914. In recommendation tasks, the overall click-through rate reached 87.1%, and forecasting accuracy remained stable between 92.4% and 90.9% across business stages, demonstrating robustness to temporal regime changes. Attention-weight analyses further provided interpretability by revealing dominant behavioral windows and influential features. These results indicated that explicitly modeling neural network dynamics—through sequence encoders, temporal attention, and adaptive updating—enhanced prediction accuracy, recommendation effectiveness, and stability under non-stationary conditions, offering a practical and scalable pathway for personalized e-commerce marketing.