| 摘 要: 为解决传统餐饮推荐服务中个性化不足、情境感知能力有限、传统推荐方法知识融合有限等问题,提出一种基于 Transformer架构的融合检索增强生成的多源感知智慧餐饮推荐算法。该算法引入分层动态秩 LoRA(Low-RankAdaptation)微调方法,根据 Transformer各层语义功能差异,自适应地调整低秩矩阵容量,显著提高了模型参数利用效率与微调灵活性;同时,构建场景多源感知与检索增强信息的上下文融合机制,通过构建显式场景提示词并结合外部知识库进行上下文融合,增强算法对复杂餐饮场景的理解与个性化推荐能力。实验结果表明,算法在训练阶段收敛速度快且loss和梯度范数均能有效降低并保持稳定。消融实验结果验证了分层动态秩 LoRA 和检索增强生成策略分别以及联合使用时的有效性,尤其在两种策略联合使用时,F1分数与余弦相似度分别提升至0.522和0.908,推荐质量显著增强。该算法在推荐准确性高、情境适配能力强、参数开销低和训练效率高等方面表现突出,具备良好的可扩展性和实际应用价值,为智慧餐饮推荐系统的智能化发展提供了一种高效、精准的解决方案 |
| 关键词: 大语言模型 参数微调 检索增强生成 推荐算法 |
|
中图分类号: TP182
文献标识码: A
|
| 基金项目: 国家重点研发计划资助(2016YFD0600101) |
|
| Multi-Source Perception Fusion and Retrieval-Augmented Generation Smart Dining Recommendation Algorithm Based on Transformer Architecture |
|
MA Chenkai, YE Ning
|
(Collage of Information Science and Technology & Artificial Intelligence, Nanjing Forestry University, Nanjing 210037, China)
machenkai@njfu.edu.cn; yening@njfu.edu.cn
|
| Abstract: To address the issues of insufficient personalization, limited contextual awareness, and constrained knowledge integration in traditional restaurant recommendation services, this paper proposes a Mult-i Source Perception Fusion and Retrieva-l Augmented Generation Smart Dining Recommendation Algorithm Based on Transformer Architecture. The algorithm introduces a hierarchical dynamic rank LoRA fine-tuning method, which adaptively adjusts the low-rank matrix capacity based on the semantic functional differences of each Transformer layer, significantly improving the model parameter utilization efficiency and fine-tuning flexibility. Additionally, a context fusion mechanism for mult-i source perceptual and retrieva-l augmented information is constructed. By designing explicit scene prompts and incorporating external knowledge bases for context fusion, the algorithm enhances its understanding of complex restaurant scenarios and improves its ability to deliver personalized recommendations. Experimental results show that the algorithm converges quickly during training, with both the loss and gradient norms effectively reduced and kept stable Retrieva-l Augmented Generation. Ablation experiments validate the effectiveness of the hierarchical
dynamic rank LoRA and retrieva-l augmented generation strategies, both individually and in combination. Particularly when both strategies are used together, the F1 score and cosine similarity are increased to 0.522 and 0.908, respectively,significantly enhancing the recommendation quality. The algorithm demonstrates outstanding performance in terms of recommendation accuracy, contextual adaptation, low parameter overhead, and high training efficiency, with excellent scalability and practical application value, offering an efficient and precise solution for the intelligent development of smart restaurant recommendation systems. |
| Keywords: large language model parameter fine-tuning Retrieva-l Augmented Generation recommendation algorithm |
