电商场景下融合语义信息与对比学习的协同过滤推荐方法

A Collaborative Recommendation Filtering Method Integrating Semantic Information and Contrastive Learning in E-Commerce Scenarios

摘要: 针对电商场景中用户–物品交互数据稀疏、协同过滤模型难以刻画物品语义内涵的问题，本文提出一种融合大语言模型语义增强与对比对齐机制的推荐模型LLMRec。该模型以LightGCN为基础框架，引入预训练语言模型生成的物品语义嵌入，并设计Top-p Gate融合机制对语义特征进行自适应激活，同时构建对比学习约束协同嵌入与语义嵌入之间的一致性。实验结果表明，在Amazon Books数据集上，LLMRec在Recall@20与NDCG@50指标上取得较优结果，其中对比对齐模块更倾向于改善更深排序范围内的整体排序结构(NDCG@50)，而语义增强模块更偏向于提升Top-K精排阶段的相关性表现。上述结果体现了两模块在不同排序深度上的互补作用，并表明该框架在电商推荐场景中具有一定的应用潜力。

Abstract: In e-commerce scenarios, user-item interaction data are usually sparse, making it difficult for conventional collaborative filtering models to effectively capture the semantic characteristics of items. To address this issue, this paper proposes a recommendation model named LLMRec, which integrates large language model-based semantic enhancement with a contrastive alignment mechanism. Built upon the LightGCN framework, LLMRec incorporates semantic embeddings generated by a pretrained language model to enrich item representations. A Top-p Gate fusion mechanism is further designed to adaptively activate semantic features, while a contrastive learning objective is introduced to enforce representation consistency between collaborative embeddings and semantic embeddings. Experimental results on the Amazon Books dataset demonstrate that LLMRec achieves superior performance in terms of Recall@20 and NDCG@50. Specifically, the contrastive alignment module tends to improve the overall ranking structure in deeper ranking ranges (NDCG@50), whereas the semantic enhancement module mainly contributes to the relevance performance in the Top-K refinement stage. These results indicate the complementary effects of the two modules at different ranking depths and suggest that the proposed framework has promising potential for e-commerce recommendation scenarios.