Generative Long-term User Interest Modeling for CTR Prediction
A new study proposes a generative approach to model long-term user interests for click-through rate prediction, addressing limitations in current two-stage frameworks. The proposed method, GenLI, uses a generative retrieval mechanism to capture latent user interests beyond target-centered behaviors, improving efficiency and accuracy in advertising and recommendation systems.
Key facts
- arXiv:2605.15905
- Modeling long-term user interests enhances CTR prediction
- Two-stage framework: GSU and ESU
- Target-centered GSU ignores latent user interests
- Matching-based retrieval is time-consuming and overlooks interaction information
- Proposed method: Generative Long-term user Interest (GenLI)
- Generative retrieval mechanism captures latent interests
- Improves efficiency and accuracy
Entities
—