Semantic User-Advertisement Alignment via Large Language Models for Personalized Recommendation

Abstract

This paper addresses the limitations of user interest modeling and semantic matching in personalized advertising recommendation by proposing a framework based on large language models. The method first embeds user behavior sequences and advertisement features, mapping discrete and continuous inputs into a unified semantic space to obtain denser and more expressive representations. A multi-head self-attention mechanism is then introduced to capture deep dependencies between users and advertisements, enabling the identification of key interest signals in long behavior sequences. A joint representation layer aligns user and advertisement representations, and cosine similarity is used to measure their matching degree, ensuring that recommendations better reflect real user needs. The entire model is trained end-to-end with a cross-entropy optimization objective to improve ranking quality and click prediction accuracy. Experiments conducted on a public advertising dataset compare the proposed approach with several mainstream methods, using NDCG, MRR, Recall@10, and Precision@10 as evaluation metrics. Results show that the proposed method achieves the best performance on all metrics. Sensitivity experiments further verify the impact of hyperparameter configurations, especially in attention head number, hidden dimension, and sequence window length, demonstrating stability and adaptability under different conditions. These findings confirm that the proposed framework improves performance across multiple metrics and exhibits strong robustness and practical value in complex and dynamic advertising recommendation scenarios.