Unified Semantic Matching with Logistic Similarity Learning for Cold-Start Advertising Recommendation

Abstract

This study addresses the cold-start problem in advertising recommendation systems and proposes an efficient modeling framework based on a unified semantic space. The research first analyzes the challenges caused by the lack of historical interaction data for new users and new ads, and maps user attributes and ad features into a shared representation space through feature encoding and latent semantic alignment to improve robustness and accuracy of matching. In the modeling process, a logistic function is used to probabilistically model similarity, and optimization is achieved by combining cross-entropy and regularization objectives to ensure stable predictive performance under sparse features and limited data. To comprehensively verify the effectiveness of the proposed method, multiple sensitivity experiments are designed, covering key factors such as learning rate, cold-start ratio, feature sparsity, and label noise rate. The experimental results show that the method maintains high levels of Precision@k, NDCG, Hit Rate, and F1-score, and effectively mitigates performance degradation caused by cold-start conditions. Overall, this work not only introduces a new modeling approach but also demonstrates through systematic experiments its practical value and adaptability in advertising cold-start recommendation scenarios.