基于大语言模型多轮对话的推荐模型研究

doi:10.3778/j.issn.1673-9418.2407087

摘要/Abstract

摘要： 近来结合大语言模型的推荐方法在提高推荐准确度和增强用户体验等方面展现出明显的优越性。然而这些方法存在没有充分利用用户信息、仅使用单轮对话无法学习用户多次交互的行为特征、大语言模型与推荐系统之间存在巨大的语义差异等问题。针对这些问题，提出了一个基于大语言模型多轮对话模式的推荐模型。该模型利用矢量量化技术将用户信息转化为用户索引，并通过微调任务把大语言模型的语言语义与推荐系统的协作语义整合，不仅学习了用户特征而且缓解了语义差异问题；将用户索引与历史交互数据拼接成提示语，再经过多轮对话机制进行推荐微调，从而学习用户交互行为之间的特征。模型在亚马逊Instructions、Arts和Games三个数据集上进行实验，结果表明模型在命中率（HR）和归一化折损累计增益（NDCG）两个评价指标上优于对比基线算法，在三个数据集上与最优对比基线算法相比，HR平均提升10.53%，NDCG平均提升5.10%，证明了模型的有效性。

关键词: 推荐系统, 序列推荐, 大语言模型, 多轮对话机制

Abstract: Recently, the recommendation method combined with large language model has shown obvious advantages in improving recommendation accuracy and enhancing user experience. However, these methods do not make full use of user information, and cannot learn the behavioral characteristics of multiple user interactions by using only a single round of dialogue, and there are huge semantic differences between large language models and recommender systems. In order to solve these problems, this paper proposes a recommendation model based on the multi-round dialogue pattern of large language model. The model uses vector quantization technology to convert user information into user indices, and integrates the language semantics of the large language model with the cooperative semantics of the recommender system through fine-tuning task, which not only learns the user characteristics but also alleviates the problem of semantic differences. The user index and historical interaction data are spliced into prompts, and then recommendations are fine-tuned through multi- round dialogue mechanism to learn the characteristics between user interaction behaviors. Experiments results on Amazon’s three benchmark datasets, such as Instructions, Arts and Games, show that the model is better than the comparison baseline algorithm in the two evaluation indices of hit rate (HR) and normalized discounted cumulative gain (NDCG). Compared with the optimal comparison baseline algorithm on the three datasets, the average increase in HR is 10.53% and the average increase in NDCG is 5.10%, which proves the effectiveness of the model.

Key words: recommendation system, sequential recommendation, large language models, multi-round dialogue mechanism

常保发, 车超, 梁艳. 基于大语言模型多轮对话的推荐模型研究[J]. 计算机科学与探索, 2025, 19(2): 385-395.

CHANG Baofa, CHE Chao, LIANG Yan. Research on Recommendation Model Based on Multi-round Dialogue of Large Language Model[J]. Journal of Frontiers of Computer Science and Technology, 2025, 19(2): 385-395.

参考文献

[1] WANG S J, HU L, WANG Y, et al. Sequential recommender systems: challenges, progress and prospects[C]//Proceedings of the 28th International Joint Conference on Artificial Intelligence, 2019: 286-299.
[2] BROWN T, MANN B, RYDER N, et al. Language models are few-shot learners[C]//Advances in Neural Information Processing Systems 33, 2020: 1877-1901.
[3] TOUVRON H, LAVRIL T, IZACARD G, et al. LLaMA: open and efficient foundation language models[EB/OL].[2024-05-13]. https://arxiv.org/abs/2302.13971.
[4] DAI S H, SHAO N L, ZHAO H Y, et al. Uncovering ChatGPT’s capabilities in recommender systems[C]//Proceedings of the 17th ACM Conference on Recommender Systems. New York: ACM, 2023: 1126-1132.
[5] LIU J, LIU C, LV R, et al. Is ChatGPT a good recommender? A preliminary study[C]//Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. New York: ACM, 2023: 1-11.
[6] GENG S J, LIU S C, FU Z H, et al. Recommendation as language processing (RLP): a unified pretrain, personalized prompt & predict paradigm (P5)[C]//Proceedings of the 16th ACM Conference on Recommender Systems. New York: ACM, 2022: 299-315.
[7] HUA W Y, XU S Y, GE Y Q, et al. How to index item IDs for recommendation foundation models[C]//Proceedings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region. New York: ACM, 2023: 195-204.
[8] LI J C, WANG M, LI J, et al. Text is all you need: learning language representations for sequential recommendation[C]//Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. New York: ACM, 2023: 1258-1267.
[9] BAO K Q, ZHANG J Z, ZHANG Y, et al. TALLRec: an effective and efficient tuning framework to align large language model with recommendation[C]//Proceedings of the 17th ACM Conference on Recommender Systems. New York: ACM, 2023: 1007-1014.
[10] HOU Y P, HE Z K, MCAULEY J, et al. Learning vector-quantized item representation for transferable sequential recommenders[C]//Proceedings of the ACM Web Conference 2023. New York: ACM, 2023: 1162-1171.
[11] LIAO J Y, LI S H, YANG Z Y, et al. LLaRA: large language-recommendation assistant[C]//Proceedings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. New York: ACM, 2024: 1785-1795.
[12] ZHENG B W, HOU Y P, LU H Y, et al. Adapting large language models by integrating collaborative semantics for recommendation[C]//Proceedings of the 2024 IEEE 40th International Conference on Data Engineering. Piscataway: IEEE, 2024: 1435-1448.
[13]    ALAYRAC J B, DONAHUE J, LUC P, et al. Flamingo: a visual language model for few-shot learning[C]//Advances in Neural Information Processing Systems 35, New Orleans, Nov 28-Dec 9, 2022: 23716-23736.
[14] HE R N, MCAULEY J. Fusing similarity models with Markov chains for sparse sequential recommendation[C]//Proceedings of the 2016 IEEE 16th International Conference on Data Mining. Piscataway: IEEE, 2016: 191-200.
[15] TANG J X, WANG K, TANG J X, et al. Personalized top-N sequential recommendation via convolutional sequence embedding[C]//Proceedings of the 11th ACM International Conference on Web Search and Data Mining. New York: ACM, 2018: 565-573.
[16] RENDLE S, FREUDENTHALER C, SCHMIDT-THIEME L, et al. Factorizing personalized Markov chains for next-basket recommendation[C]//Proceedings of the 19th International Conference on World Wide Web. New York: ACM, 2010: 811-820.
[17] HIDASI B, KARATZOGLOU A, BALTRUNAS L, et al. Session-based recommendations with recurrent neural networks[EB/OL]. [2024-05-13]. https://arxiv.org/abs/1511.06939.
[18]    WANG C, NIEPERT M. State-regularized recurrent neural networks[C]//Proceedings of the 36th International Conference on Machine Learning, 2019: 6596-6606.
[19] KANG W C, MCAULEY J. Self-attentive sequential recommendation[C]//Proceedings of the 2018 IEEE International Conference on Data Mining. Piscataway: IEEE, 2018: 197-206.
[20]    VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[C]//Advances in Neural Information Processing Systems 30, Long Beach, Dec 4-9, 2017: 5998-6008.
[21]    SUN F, LIU J, WU J, et al. BERT4Rec: sequential recommendation with bidirectional encoder representations from transformer[C]//Proceedings of the 28th ACM International Conference on Information and Knowledge Management. New York: ACM, 2019: 1441-1450.
[22] ZHOU K, WANG H, ZHAO W X, et al. S3-Rec: self-supervised learning for sequential recommendation with mutual information maximization[C]//Proceedings of the 29th ACM International Conference on Information and Knowledge Management. New York: ACM, 2020: 1893-1902.
[23] ZHOU K, YU H, ZHAO W X, et al. Filter-enhanced MLP is all you need for sequential recommendation[C]//Proceedings of the ACM Web Conference 2022. New York: ACM, 2022: 2388-2399.
[24] LIU Z H, MEI S, XIONG C Y, et al. Text matching improves sequential recommendation by reducing popularity biases[C]//Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. New York: ACM, 2023: 1534-1544.
[25] LEE A N, HUNTER C J, RUIZ N. Platypus: quick, cheap, and powerful refinement of LLMs[C]//Advances in Neural Information Processing Systems 36, New Orleans, Dec 10- 16, 2023: 321-332.
[26] CUI Z, MA J, ZHOU C, et al. M6-Rec: generative pretrained language models are open-ended recommender systems[C]//Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. New York: ACM, 2022: 1472-1482.
[27] ZHU Y C, WU L, GUO Q, et al. Collaborative large language model for recommender systems[C]//Proceedings of the ACM Web Conference 2024. New York: ACM, 2024: 3162-3172.
[28] EVURU C K, GHOSH S, KUMAR S, et al. CoDa: constrained generation based data augmentation for low-resource NLP[C]//Proceedings of the Findings of the Association for Computational Linguistics: NAACL 2024. Stroudsburg: ACL, 2024: 425-436.
[29]    RAJPUT S, MEHTA N, SINGH A, et al. Recommender systems with generative retrieval[C]//Advances in Neural Information Processing Systems 36, New Orleans, Dec 10-16, 2023: 452-463.
[30] VAN DEN OORD A, VINYALS O. Neural discrete representation learning[C]//Advances in Neural Information Processing Systems 30, Long Beach, Dec 4-9, 2017: 271-284.
[31] LIU Z C, CHENG K T, HUANG D, et al. Nonuniform-to-uniform quantization: towards accurate quantization via generalized straight-through estimation[C]//Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2022: 4932-4942.
[32] MA C, KANG P, LIU X, et al. Hierarchical gating networks for sequential recommendation[C]//Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York: ACM, 2019: 825-833.
[33] ZHANG T T, ZHAO P P, LIU Y C, et al. Feature-level deeper self-attention network for sequential recommendation[C]//Proceedings of the 28th International Joint Conference on Artificial Intelligence. New York: ACM, 2019: 4320-4326.
[34] ZHANG Z, WANG B. Prompt learning for news recommendation[C]//Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. New York: ACM, 2023: 227-237.