Generative Long-term User Interest Modeling for Click-Through Rate Prediction
The paper presents a new model called GenLI for improving click-through rate (CTR) prediction by modeling long-term user interests. It addresses limitations of existing methods by incorporating interaction information among user behaviors and reducing time complexity in behavior retrieval. The proposed framework enhances the diversity of user interests and balances accuracy with efficiency in CTR prediction.
- ▪GenLI consists of three modules: interest generation, behavior retrieval, and interest fusion.
- ▪The interest generation module creates multiple interest distributions that are target-independent.
- ▪The behavior retrieval module simplifies the selection of related behaviors to reduce time complexity.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Information Retrieval arXiv:2605.15905 (cs) [Submitted on 15 May 2026] Title:Generative Long-term User Interest Modeling for Click-Through Rate Prediction Authors:Jiangli Shao, Kaifu Zheng, Hao Fang, Huimu Ye, Zhiwei Liu, Bo Zhang, Shu Han, Xingxing Wang View a PDF of the paper titled Generative Long-term User Interest Modeling for Click-Through Rate Prediction, by Jiangli Shao and 7 other authors View PDF HTML (experimental) Abstract:Modeling long-term user interests with massive historical user behaviors enhances click-through rate (CTR) prediction performance in advertising and recommendation systems.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
