Generative Long-term User Interest Modeling for Click-Through Rate Prediction DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Information systems Recommender systems CCS: Information systems Computational advertising
Jiangli Shao, Kaifu Zheng, Hao Fang, Huimu Ye, Zhiwei Liu, Bo Zhang, Shu Han, Xingxing Wang · 2026
Read Generative Long-term User Interest Modeling for Click-Through Rate Prediction DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Information systems Recommender systems CCS: Information systems Computational advertising by Jiangli Shao, Kaifu Zheng, Hao Fang, Huimu Ye, Zhiwei Liu, Bo Zhang, Shu Han, Xingxing Wang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-15905.txt.
Introduction
Jiangli Shao and seven colleagues propose GenLI, a generative model of long-term user interests for click-through rate prediction. Two-step systems first retrieve top-k past acts matching a target item, then build features. But target-centered retrieval misses latent tastes, and pairwise matching grows costly as histories grow. GenLI generates interest distributions independent of the target, folding in links among past acts. It picks past acts by lookup in O(1) time, and fuses tastes through gating. The design improves taste diversity and balances accuracy against efficiency.
Catalog record
| CREATOR | Jiangli Shao, Kaifu Zheng, Hao Fang, Huimu Ye, Zhiwei Liu, Bo Zhang, Shu Han, Xingxing Wang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48280 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.15905
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Generative Long-term User Interest Modeling for Click-Through Rate Prediction DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Information systems Recommender systems CCS: Information systems Computational advertising Jiangli Shao, Kaifu Zheng, Hao Fang, Huimu Ye, Zhiwei Liu, Bo Zhang, Shu Han, Xingxing Wang Abstract . Modeling long-term user interests with massive historical user behaviors enhances click-through rate (CTR) prediction performance in advertising and recommendation systems. Typically, a two-stage framework is widely adopted, where a general search unit (GSU) first retrieves top- k k relevant behaviors towards the target item, and an exact search unit (ESU) generates interest features via tailored attention. However, current target-centered GSU would ignore other latent user interests, leading to incomplete and biased interest features. Additionally, the matching-based retrieval process in GSUs depends on the pairwise similarity score between target item and each historical behavior, which not only becomes time-consuming for online services as user behaviors continue to grow, but also overlooks the interaction information among user behaviors. To combat these problems, we propose a Gen erative L ong-term user I nterest model named GenLI for CTR prediction. GenLI consists of an interest generation module (IGM), a behavior retrieval module (BRM), and an interest fusion module (IFM). The IGM generates multiple interest distributions to indicate different aspects of real-time user interests, which is target-independent and incorporates interaction information among behaviors, ensuring complete and diverse interest features. The BRM selects related behaviors via a simple lookup operation, reducing the time complexity for weighting each behavior to O ( 1 ) O(1) . Finally, the IFM uses delicate gating mechanisms to generate interest features. Based on the generation process, GenLI improves the diversity of user interests and avoids complex matching-based behavioral retrieval, achieving a better balance between accuracy and efficiency for CTR prediction. Extensive offline and online experiments demonstrate the superior performance of GenLI. Notably, GenLI has been deployed in a real-world platform, serving the main traffic of hundreds of millions of users, showing its application value. 1. Introduction Click-through rate (CTR) prediction is a core task for online commercial platforms ( 14 ; 31 ) . In recommendation systems, accurate CTR helps deliver the most appropriate content to users, enhancing user satisfaction while increasing user engagement ( 5 ; 28 ) . In advertising systems, CTR influences the calculation of Effective Cost Per Mille (eCPM) in the cost-per-click (CPC) billing model, thus directly affecting ad revenue ( 3 ; 8 ) . Hence, it is crucial for platforms to build an accurate CTR prediction model. The core of CTR modeling is personalization ( 13 ; 29 ; 35 ) . To accomplish this, CTR models should be highly individualized, leveraging user behavioral data to its full potential. These historical user behaviors contain user interest patterns ( 32 ; 21 ) , usually serving as valid features for CTR prediction. Therefore, user interest modeling has become a hot topic in both industry and academia. Generally, user interest modeling includes capturing short-term and long-term user interests from user behavior sequences. For short-term user interest modeling, researchers aggregate users’ short-term behaviors with fine-grained attention mechanisms, such as DIN ( 37 ) , DIEN ( 36 ) , DSIN ( 7 ) , etc. In long-term user interest modeling, a cascaded two-stage framework is widely adopted to balance the accuracy and inference latency ( 20 ; 22 ) . In this framework, a GSU first retrieves top- k relevant behaviors towards the target item with simple similarity metrics such as BM25 ( 22 ) , inner product ( 20 ) , or hash fingerprint ( 1 ; 4 ) . Subsequently, based on more complicated attention mechanisms, an ESU aggregates retrieved behaviors to generate long-term user interest features. Compared with short-term interests, long-term interests provide richer information on user interest patterns and display a more complete view of user preference, leading to better CTR prediction performance in systems. Meanwhile, with the development of platforms, more user behavior data are gradually accumulated, supporting the modeling of interests over a longer range. Thus, more efforts have been made to model long-term user interests in recent years. However, current long-term interest modeling approaches have some deficiencies that may impair performance. First, existing GSUs are target-centered, considering only target-related interests while ignoring other latent user interests. Users generally have multiple interests ( 34 ) , and focus on just one aspect can introduce bias in interest modeling. In addition, the retrieval process in GSUs depends on the similarity score between historical behaviors and target items. This score is calculated using a matching-based paradigm, requiring comparing the target item with each historical behavior. This time-consuming process will become a bottleneck when user behaviors extend, limiting the length of user behavior sequences that can be considered in online services. Moreover, GSUs adopt a pairwise architecture that compares the target item with every single behavior, ignoring the interaction information regarding user behaviors. This results in inaccurate similarity metrics for GSUs, further degrading the CTR prediction performance. To address these issues, we propose a generative long-term user interest modeling framework named GenLI for CTR prediction. Unlike existing methods that rely on target-centered retrieval, GenLI uses a generative approach to directly produce diverse and comprehensive user interest distributions and selects behaviors via distribution lookup to improve efficiency. In detail, GenLI is composed of an interest generation module (IGM), a behavior retrieval module (BRM), and an interest fusion module (IFM). The IGM takes the recent few behaviors as input and outputs the implicit, explicit, and relative interest distributions. These distributions collectively encapsulate the real-time user interests. Importantly, this module operates independently of any target item, ensuring that the generated distributions encompass a broader range of interests beyond just the target items. Based on each interest distribution, the BRM selects a subset of behaviors that best fit the corresponding type of interest through lookup scores. The scores describe the importance of behaviors within multiple interest distributions, reflecting diversified interests. In addition, the score incorporates interaction information within behaviors, which is more precise than simple pairwise metrics. Otherwise, the lookup process is more efficient as it avoids the comparison between the target item and each historical behavior, reducing the time complexity to weight each behavior to O(1) . Finally, the IFM separately aggregates the selected behaviors from different interest distributions and eventually uses gating mechanisms to produce long-term user interest features, serving as the main features for CTR prediction. Relying on the generative method, GenLI improves the diversity of user interests and avoids complex matching-based behavioral retrieval, achieving a better balance between accuracy and efficiency. GenLI demonstrates superior performance in offline experiments and achieves a promotion of 1.56% in RPM (Revenue per Mille) in online A/B testing. As its application value, GenLI has been deployed in online advertising systems in the real-world platform and serves hundreds of millions of customers every day. The main contributions of our work are summarized as follows: • To the best of our knowledge, we are the first to introduce generative methods in modeling long-term user interests for CTR prediction tasks. • We propose a framework named GenLI for long-term user interest modeling,