# A Unified Framework for Multi-Domain CTR Prediction via Large Language Models

**Authors:** Zichuan Fu, Xiangyang Li, Chuhan Wu, Yichao Wang, Kuicai Dong, Xiangyu Zhao, Mengchen Zhao, Huifeng Guo, Ruiming Tang  
**arXiv:** [2312.10743](https://arxiv.org/abs/2312.10743)  
**Date:** 2023-12-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2312.10743v1/

## Key points

- Click-through rate (CTR) prediction in online recommendation platforms is important
- Challenges associated with multi-domain CTR prediction
- Traditional models using discrete identifiers limit generalization and result in performance drops in certain domains
- Proposed solution called Uni-CTR utilizes a Large Language Model (LLM) to capture commonalities between domains and domain-specific networks for unique characteristics
- Masked loss strategy decouples domain-specific networks from the backbone LLM, allowing flexibility and scalability
- Experimental results show Uni-CTR outperforms state-of-the-art MDCTR models significantly and demonstrates effectiveness in zero-shot prediction
- Uni-CTR applied in industrial scenarios confirms its efficiency
- Unified framework for multi-domain CTR prediction leveraging semantic representations and domain-specific networks
- Uni-CTR model showcases superior performance compared to existing models and proves effective in real-world applications.

## Summary

, , , , 

The importance of click-through rate (CTR) prediction in online recommendation platforms is discussed in this paper, along with the challenges associated with multi-domain CTR prediction. Traditional models often use discrete identifiers to represent domains, limiting their ability to generalize and resulting in performance drops in certain domains. To overcome these limitations, the authors propose a novel solution called Uni-CTR. This model utilizes a backbone Large Language Model (LLM) to learn layer-wise semantic representations that capture commonalities between domains, while also incorporating domain-specific networks to capture unique characteristics of each domain. A masked loss strategy is designed to decouple the domain-specific networks from the backbone LLM, allowing for flexibility and scalability when incorporating new or removing existing domains. Experimental results on three public datasets show that Uni-CTR outperforms state-of-the-art MDCTR models significantly and demonstrates remarkable effectiveness in zero-shot prediction. The authors have also applied Uni-CTR in industrial scenarios and confirmed its efficiency. In summary, this paper presents a unified framework for multi-domain CTR prediction that addresses the limitations of traditional models by leveraging semantic representations and domain-specific networks. The proposed Uni-CTR model showcases superior performance compared to existing models and proves effective in real-world applications.