GenCI: Generative Modeling of User Interest Shift via Cohort-based Intent Learning for CTR Prediction Conference: Proceedings of the ACM Web Conference 2026; April 13–17, 2026; Dubai, United Arab Emirates Proceedings of the ACM Web Conference 2026 (WWW ’26), April 13–17, 2026, Dubai, United Arab Emi
Kesha Ou, Zhen Tian, Wayne Xin Zhao 🖂, Hongyu Lu, Ji-Rong Wen · 2026
Read GenCI: Generative Modeling of User Interest Shift via Cohort-based Intent Learning for CTR Prediction Conference: Proceedings of the ACM Web Conference 2026; April 13–17, 2026; Dubai, United Arab Emirates Proceedings of the ACM Web Conference 2026 (WWW ’26), April 13–17, 2026, Dubai, United Arab Emi by Kesha Ou, Zhen Tian, Wayne Xin Zhao 🖂, Hongyu Lu, Ji-Rong Wen (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-2601-18251.txt.
Introduction
Kesha Ou and co-authors model how a user's interests shift over time, for better click prediction. Their 2026 preprint presents GenCI. Click-through rate prediction is pivotal in online ads and recommenders. Most models match a candidate item against the user's history. That matching can overfit to the traits that dominated the past. It then fails when interests move fast. The authors instead model interest shift in a generative way, learning intent through cohorts of similar users.
Cohort views let the model borrow strength from people with like patterns. A user whose taste is turning can be read against a cohort that already turned. Fresher intent reads should lift both ads and feeds.
Catalog record
| CREATOR | Kesha Ou, Zhen Tian, Wayne Xin Zhao 🖂, Hongyu Lu, Ji-Rong Wen |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-18251
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
GenCI: Generative Modeling of User Interest Shift via Cohort-based Intent Learning for CTR Prediction Conference: Proceedings of the ACM Web Conference 2026; April 13–17, 2026; Dubai, United Arab Emirates Proceedings of the ACM Web Conference 2026 (WWW ’26), April 13–17, 2026, Dubai, United Arab Emirates DOI: 10.1145/3774904.3792612 ISBN: 979-8-4007-2307-0/2026/04 Thanks: 🖂 Corresponding author. CCS: Information systems Recommender systems Kesha Ou, Zhen Tian, Wayne Xin Zhao 🖂, Hongyu Lu, Ji-Rong Wen Abstract . Click-through rate (CTR) prediction plays a pivotal role in online advertising and recommender systems. Despite notable progress in modeling user preferences from historical behaviors, two key challenges persist. First, exsiting discriminative paradigms focus on matching candidates to user history, often overfitting to historically dominant features and failing to adapt to rapid interest shifts. Second, a critical information chasm emerges from the point-wise ranking paradigm. By scoring each candidate in isolation, CTR models discard the rich contextual signal implied by the recalled set as a whole, leading to a misalignment where long-term preferences often override the user’s immediate, evolving intent. To address these issues, we propose GenCI , a generative user intent framework that leverages semantic interest cohorts to model dynamic user preferences for CTR prediction. The framework first employs a generative model, trained with a next-item prediction (NTP) objective, to proactively produce candidate interest cohorts. These cohorts serve as explicit, candidate-agnostic representations of a user’s immediate intent. A hierarchical candidate-aware network then injects this rich contextual signal into the ranking stage, refining them with cross-attention to align with both user history and the target item. The entire model is trained end-to-end, creating a more aligned and effective CTR prediction pipeline. Extensive experiments on three widely used datasets demonstrate the effectiveness of our approach. † † cc-license: by Resource Availability: The source code of this paper has been made publicly available at https://doi.org/10.5281/zenodo.18344901 . 1. Introduction The click-through rate (CTR) prediction task, which aims to estimate the probability of a user clicking an item, is essential for web applications like recommender systems to improve user experience and platform revenue. Various models ( 49 ; 24 ; 48 ) have been proposed to capture user interests from historical behaviors. Despite the notable achievements, two fundamental challenges persist in contemporary paradigms. First, existing models have limited ability to capture shifts in user interest. Prior studies ( 47 ; 49 ; 48 ) demonstrate that leveraging behavioral histories can improve performance, primarily by using attention mechanisms to weigh the relevance of past behaviors with respect to the target item. However, these models are designed to assess candidate relevance rather than construct a standalone representation of the user’s current intent, making it difficult to handle evolving interests. As illustrated in the upper part of Figure 1 , when a user’s focus shifts from iPhone accessories to MacBook accessories , such models tend to assign a high score to an iPhone screen protector due to the abundance of Apple-related items in the user’s history, thereby failing to capture the evolving preference. Moreover, the fundamentally discriminative nature of these approaches often induces shortcut learning ( 41 ; 10 ; 45 ) , where the model relies on local strongly predictive historical features instead of capturing the full spectrum of user interests, further hindering its ability to model dynamic user preferences. Second, the point-wise CTR prediction paradigm evaluates each recalled candidate in isolation, ignoring the latent interest signal implied by the candidate set as a whole. In modern recall-then-rank pipelines, this context-agnostic approach introduces a substantial information gap. The recall module retrieves a candidate set that reflects hypotheses about the user’s immediate intent, forming an contextual interest cohort( i.e., a group of semantically similar items). However, this rich contextual signal is later discarded under the point-wise ranking schema, leading to misalignment between the two stages. As illustrated in the lower part of Figure 1 , when a user shifts their focus from history to photography, relevant photography books may be successfully recalled, yet the context-agnostic ranker might still prioritize familiar history items driven by strong long-term preferences, failing to capture the user’s evolving intent. To overcome these limitations, we introduce a generative paradigm to more effectively model user interest shifts through an auxiliary next-item prediction (NTP) task. Our approach is grounded in modeling the user’s chronologically ordered click sequence, through which the causal decision-making process and evolving intent can be inferred ( 20 ) . Based on this foundation, our model proactively generates semantic IDs that represent a hypothesis about the user’s next interaction, overcoming the limitation of discriminative methods that passively evaluate target relevance without forming a candidate-agnostic representation of current intent. Furthermore, because this generative prediction is made over the entire item space ( 29 ) , the resulting semantic IDs encode globally coherent information about user interests, mitigating the overfitting of discriminative methods to local high-frequency features. By modeling the interaction between these semantic IDs and the target item, we inject this rich contextual foundation into the CTR prediction stage, enabling the ranking model to be aware of the recall-stage context ( i.e., why an item was retrieved alongside others). This design addresses the inherent limitation of point-wise CTR models, which evaluate items independently without awareness of the retrieval context, thereby achieving better alignment between the recall and ranking stages. In this paper, we propose GenCI , a novel generative user intent framework that constructs semantic interest cohorts as explicit intent representations for CTR prediction. First, hierarchical quantization is applied to organize items into semantically coherent cohorts, forming latent interest groups that capture fine-grained user preferences. Based on this representation, we employ a generative sequential model built upon a Transformer architecture to perform the NTP task, which extracts dual-aspect user intent: the encoder produces a general long-term interest profile, while the decoder generates candidate interest cohorts that capture the user’s immediate, short-term intent. These candidate cohorts are then incorporated as dynamic contextual signals in the ranking stage. To mitigate the noise introduced by items selected merely due to coarse semantic similarity, we design a hierarchical candidate-aware modeling module that refines the cohorts’ intent via cross-attention, aligning them with both historical behaviors and the target item to produce highly personalized and dynamic interest representations. Finally, the multi-perspective intent signals are integrated with conventional features and fed into MLPs for CTR estimation, with the entire model trained end-to-end under a joint optimization scheme, augmented by a self-supervised regularization that stabilizes learning and improves robustness. The contributions of this paper are summarized as follows: \bullet We propose GenCI , a novel framework that introduces a generative paradigm for modeling dynamic user intent. It leverages a NTP task to proactively generate semantically coherent interest cohorts, enabling the capture of multi-faceted user preferences. \bullet We introduce a hierarchical candidate-aware network that personalizes the generated cohorts via cross-attention, thereby injecting them as dynamic signals to make the point-wise ranker aware of the recall signal. \bullet We conduct extensive experiments on three widely used datasets. GenCI consistently outperforms a number of competitive baselines, showing the effectiveness of our model. Figure 1 . Illustration of two key challenges in CTR prediction. Top: Relevance-based models focus on long-term preference matching and fail to adapt to short-term intent shifts ( e.g., from iPhone