Source-linked AI summary
GenRec: Large Language Model for Generative Recommendation
Jianchao Ji, Zelong Li, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Juntao Tan, Yongfeng Zhang
TL;DR
Recommendation research has underexplored text-based generative use of LLMs, especially where models directly generate target items rather than score candidates. GenRec uses prompted textual interaction sequences to fine-tune LLaMA, and it performs significantly better than P5 on MovieLens 25M while P5 performs better on Amazon Toys.
Problem
Text-based generative recommendation with LLMs remains relatively unexplored, while many recommendation models omit textual item information.
Method
GenRec formats user-item interactions into prompted textual sequences and fine-tunes a LLaMA backbone to generate the next item.
Results
GenRec performs significantly better on MovieLens 25M, while P5 performs better on Amazon Toys.
Takeaways & Limitations
GenRec provides a foundational text-based generative recommendation framework with potential adaptability across applications and LLM backbones.
Takeaways & Limitations
Future work includes more sophisticated prompts, richer interaction data such as ratings or reviews, and evaluation with different LLMs.
Abstract
from arXiv · showhide
In recent years, large language models (LLM) have emerged as powerful tools for diverse natural language processing tasks. However, their potential for recommender systems under the generative recommendation paradigm remains relatively unexplored. This paper presents an innovative approach to recommendation systems using large language models (LLMs) based on text data. In this paper, we present a novel LLM for generative recommendation (GenRec) that utilized the expressive power of LLM to directly generate the target item to recommend, rather than calculating ranking score for each candidate item one by one as in traditional discriminative recommendation. GenRec uses LLM's understanding ability to interpret context, learn user preferences, and generate relevant recommendation. Our proposed approach leverages the vast knowledge encoded in large language models to accomplish recommendation tasks. We first we formulate specialized prompts to enhance the ability of LLM to comprehend recommendation tasks. Subsequently, we use these prompts to fine-tune the LLaMA backbone LLM on a dataset of user-item interactions, represented by textual data, to capture user preferences and item characteristics. Our research underscores the potential of LLM-based generative recommendation in revolutionizing the domain of recommendation systems and offers a foundational framework for future explorations in this field. We conduct extensive experiments on benchmark datasets, and the experiments shows that our GenRec has significant better results on large dataset.
1 INTRODUCTION
Recommendation systems face cold start, sparsity, and scalability challenges, while many models omit textual item information. GenRec addresses this gap through text-based generative recommendation with LLMs.
- Traditional recommenders struggle with cold start, sparse interactions, and scaling large user-item matrices.
- Many recommendation models use user and item indexes, omitting textual information such as item titles and categories.
- GenRec incorporates textual item information to improve generative recommendation and capture semantically analyzable features in item names.
- The paper evaluates GenRec against other LLM recommendation models and aims to encourage further research and adoption of LLM-based recommenders.
- Generative recommendation directly produces a target item instead of scoring and sorting every candidate item individually.
2 RELATED WORK
Prior recommendation research includes collaborative, content-based, hybrid, and LLM-based approaches. GenRec extends emerging text-based generative recommendation, whose potential remains incompletely explored.
- Collaborative filtering models user-item interaction patterns, while content-based systems use item features and learned content representations.
- Large language models offer recommendation systems rich semantic and long-range modeling of user-item interactions.
- P5 demonstrated that recommendation can be formulated as a natural-language task within a unified multi-task system.
- The use of large language models to understand and generate text-based recommendations has not been fully explored.
- GenRec proposes a text-based generative recommendation approach intended to address limitations of previous work.
3 METHOD
GenRec reformats interaction histories into prompted training sequences and fine-tunes an LLM to predict the next item. The framework uses LLaMA while remaining compatible with other LLMs.
- GenRec formats item interaction sequences with prompts, fine-tunes an LLM, and predicts subsequent items the user may interact with.
- LLaMA is the backbone, while the framework is designed for integration with other LLMs.
- 3.1 Sequence Generation: Each generated training sequence contains an instruction, an input interaction history excluding the latest item, and the latest interaction as output.
- 3.1 Sequence Generation: Multiple prompts are designed to improve the LLM’s comprehension of recommendation tasks during sequence generation.
- LLaMA-LoRA reduces GPU memory requirements enough to fine-tune and infer on a single 24GB GPU.
4 EXPERIMENTS
GenRec is evaluated on Amazon and MovieLens recommendation datasets using HR and NDCG. Performance varies by dataset: P5 leads on Amazon Toys, whereas GenRec performs significantly better on MovieLens 25M.
- 4.1 Dataset: The experiments use Amazon purchase histories and MovieLens ratings with associated metadata.
- 4.1 Dataset: For each user sequence, the latest interaction is held out for testing, the second latest for validation, and earlier interactions for training.
- Performance is evaluated using Hit Ratio and Normalized Discounted Cumulative Gain.
- 4.5 Performance Comparison: P5 performs better on Amazon Toys, while GenRec performs significantly better on MovieLens 25M.
- 4.5 Performance Comparison: The authors attribute GenRec’s stronger MovieLens 25M performance partly to richer interaction information and its ability to capture user interests and item content.
5 CONCLUSION
GenRec is presented as a text-based generative recommendation framework centered on semantically rich item names, with claimed personalization, contextual relevance, and broad LLM integration. The authors identify more sophisticated prompting, richer interaction data, and evaluation across different LLMs as future directions.
- 5 CONCLUSION: GenRec uses the semantic richness of item names to support more personalized and contextually relevant recommendations across applications.The framework is described as adaptable across diverse applications.
- 5 CONCLUSION: The framework is designed for integration with any large language model, broadening its potential utility.
- 5 CONCLUSION: Future work will refine prompts, incorporate ratings or reviews, and test GenRec with different LLMs to examine potential benefits and trade-offs.