Large interest network for click-through rate prediction.

Abstract

Contemporary recommender systems face two critical challenges in click-through rate (CTR) prediction. First, traditional methods are confined to learning from closed training datasets, lacking access to broader world knowledge. Second, existing approaches struggle to achieve fine-grained interest modeling while maintaining efficient inference. To address these challenges in CTR prediction, we propose Large Interest Network (LIN), a unified framework that leverages large language models (LLMs), contrastive learning, and clustering techniques. LIN utilizes LLMs to generate semantically rich profiles for users and items. This external knowledge is then incorporated through Semantic Hard Negatives Contrastive Learning that aligns different representation spaces while preserving their complementary strengths. Furthermore, we introduce Cluster-based Interest Representation using LLM profile cluster centroids, combining the efficiency of two-tower models with the accuracy of interest-modeling approaches. Extensive experiments on multiple public datasets demonstrate that LIN consistently outperforms state-of-the-art methods, achieving improvements of 0.46 %-2.75 % in AUC and 7.52 %-15.18 % in LogLoss compared to the best baselines, while delivering 36×-244× faster inference speed compared to interest-modeling approaches. Our code is publicly available at https://github.com/AllminerLab/LargeInterestNetwork.