# CoLLM: Integrating Collaborative Embeddings into Large Language Models for Recommendation

**Authors:** Yang Zhang, Fuli Feng, Jizhi Zhang, Keqin Bao, Qifan Wang, Xiangnan He  
**arXiv:** [2310.19488](https://arxiv.org/abs/2310.19488)  
**Date:** 2023-10-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.19488v1/

## Key points

- Leveraging Large Language Models as Recommenders (LLMRec) has gained significant attention in user preference modeling.
- Existing LLMRec approaches prioritize text semantics and neglect collaborative information from user-item interactions.
- Text-emphasizing approaches excel in cold-start scenarios but may yield sub-optimal performance in warm-start situations.
- CoLLM is an innovative LLMRec methodology that incorporates collaborative information into LLMs for recommendation.
- CoLLM captures collaborative information through an external traditional model and maps it to the input token embedding space of LLMs.
- This process forms collaborative embeddings that can be effectively utilized by LLMs without modifying the models themselves.
- CoLLM allows for flexible employment of various collaborative information modeling techniques.
- By leveraging external models, CoLLM ensures effective modeling of collaborative information without sacrificing the benefits provided by large language models.
- Extensive experiments validate that CoLLM enhances recommendation performance compared to text-emphasizing approaches alone.
- The code and data for CoLLM are available on GitHub at https://github.com/zyang1580/CoLLM.

## Summary

In the field of user preference modeling, Leveraging Large Language Models as Recommenders (LLMRec) has gained significant attention and introduced fresh perspectives. However, existing LLMRec approaches tend to prioritize text semantics while neglecting valuable collaborative information from user-item interactions in recommendations. While these text-emphasizing approaches excel in cold-start scenarios, they may yield sub-optimal performance in warm-start situations. To address this limitation and provide superior recommendations for both cold and warm start scenarios, the authors propose CoLLM. CoLLM is an innovative LLMRec methodology that seamlessly incorporates collaborative information into LLMs for recommendation. It captures collaborative information through an external traditional model and maps it to the input token embedding space of LLMs. This process forms collaborative embeddings that can be effectively utilized by LLMs without modifying the models themselves. The integration of collaborative information into LLMs allows for flexible employment of various collaborative information modeling techniques. By leveraging external models to capture collaborative signals and incorporating them into LLMs, CoLLM ensures effective modeling of collaborative information without sacrificing the benefits provided by large language models. Extensive experiments validate that CoLLM adeptly integrates collaborative information into LLMs, resulting in enhanced recommendation performance compared to text-emphasizing approaches alone. The authors have made the code and data for CoLLM available on GitHub at https://github.com/zyang1580/CoLLM. In summary, CoLLM presents a novel approach to recommendation systems by seamlessly integrating collaborative embeddings into large language models.