Multi-Level Attention and Sequence Modeling for Dynamic User Interest Representation in Real-Time Advertising Recommendation

Abstract

This study investigates sequence modeling and attention optimization in real-time advertising recommendation, aiming to address the limitations of traditional methods in capturing dynamic user interests. It first analyzes the temporal dependencies of user behavior in advertising scenarios and points out that models based only on static correlations cannot fully represent the dynamic changes of user preferences. To solve this, a framework integrating sequence modeling with multi-level attention is proposed, which models user actions such as clicks, browsing, and purchases in order and achieves a unified representation of short-term interests and long-term preferences. The design introduces embedding layers to enhance feature representation and applies self-attention to highlight key behavior fragments, balancing global and local feature modeling. Residual connections and regularization are further incorporated to improve stability and generalization. Experiments compare the proposed method with multiple baselines on Precision@10, Recall@10, NDCG, and AUC, while also conducting sensitivity tests on regularization strength, model update intervals, cold-start ratios, and label noise. Results demonstrate that the proposed framework achieves consistent advantages across metrics, effectively capturing complex patterns in user behavior sequences and showing strong effectiveness and adaptability in real-time advertising recommendations.