Empirical Study on the Application of Deep Learning in User Behavior Prediction and Personalized Recommendation in E-Commerce
Guoyu Diao, Chengcheng Li, Qian Liu, Zegang Liu · Journal of Organizational and End User Computing · 2025
Personalized recommendation systems are crucial for improving user experience and business performance in e-commerce. However, existing models face two major challenges: (a) an imbalance between short-term ranking accuracy and long-term engagement optimization, and (b) limited utilization of multi-modal information, resulting in suboptimal contextual understanding and poor cold-start performance. Traditional sequential models effectively capture short-term user preferences but fail to consider long-term engagement dynamics, while reinforcement learning (RL)-based approaches optimize engagement but suffer from high computational complexity and slow convergence. To address these challenges, we propose HDL-RecBERT, a hybrid recommendation framework that integrates transformer-based sequential modeling, RL for long-term optimization, adaptive multi-modal feature fusion, and contrastive self-supervised learning. The model employs self-attention mechanisms to model user behavior, RL to maximize cumulative user engagement, cross-modal attention to dynamically fuse multi-modal data, and contrastive learning to enhance cold-start recommendation performance. Extensive experiments on real-world e-commerce datasets show that HDL-RecBERT outperforms state-of-the-art baselines, achieving an 11.2% improvement in HR@10 and a 9.5% increase in NDCG@10, while RL improves cumulative reward (CR) by 10.2% and contrastive learning enhances cold-start recall by 9.3%. These results demonstrate HDL-RecBERT's ability to balance short-term and long-term optimization, improve recommendation diversity, and enhance adaptability in cold-start scenarios, making it a promising solution for next-generation recommendation systems.