**A scalable CTR model inspired by LLM to explore scaling laws **
🔗 Paper (RecSys '25) | 💻 Code

Each Unified Attention Block (UAB) contains:

• Self-Attention: Spatiotemporal behavior modeling
• Cross-Attention: User profile-guided importance scoring
• Dual Alignment Attention: Feature selection
• RMSNorm + SwiGLU FFN (LLM-inspired)

📌 Input: Target-aware sequence = User behaviors + candidates
📌 Output: P(click|S,p,c) = σ(MLP(E_block[-1,:], e_p, e_other))

Due to industrial deployment constraints, we release:

• File: ./handle_layer/handle_lib/handle_rec_unit.py
• Key classes:
  • Mix1k_SUAN: For industrial dataset
  • Eleme_SUAN: For Eleme dataset

• exp/user1/Mix1k_SUAN/: Industrial dataset config
• exp/user1/Eleme_SUAN/: Eleme dataset config

@inproceedings{lai2025exploring,
  title={Exploring Scaling Laws of CTR Model for Online Performance Improvement},
  author={Lai, Weijiang and Jin, Beihong and Zhang, Jiongyan and Zheng, Yiyuan and Dong, Jian and Cheng, Jia and Lei, Jun and Wang, Xingxing},
  booktitle={Proceedings of the Nineteenth ACM Conference on Recommender Systems},
  pages={114--123},
  year={2025},
  organization={ACM}
}

• Email: [email protected]
• Affiliation: Institute of Software, Chinese Academy of Sciences
• GitHub: https://github.com/laiweijiang/SUAN

⭐ Star us if you find it useful!