基于物品描述和评论的多粒度注意力机制的推荐  

Multi-Granularity Attention Mechanism for Recommendation Systems Based on Item Descriptions and Reviews

作　　者：刘世新 Shixin Liu(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海

出　　处：《建模与仿真》2024年第3期2429-2440,共12页Modeling and Simulation

摘　　要：本研究针对传统推荐系统依赖评分信息、面临数据稀疏性的局限,以及深度学习模型的可解释性不足,提出了一种融合评分数据和文本信息(评论、物品描述)的多粒度注意力机制的推荐模型(Multi-Grained Attention Recommendation,MGAR)。本文采用两个并行的卷积神经网络(CNN)分别处理用户和物品的评论与描述,并通过词级、短语级和句子级的注意力机制来提取不同层次的语义信息,实现评论文本的深度融合,从而更有效地捕捉用户偏好和物品特性。在Amazon上的4个子数据集的实验结果表明,本文提出的模型在预测准确度上均优于传统的基于评分的推荐模型和近年来流行的DeepCoNN模型。This study addresses the challenges faced by traditional recommendation systems that rely on rating information and contend with data sparsity,as well as the lack of interpretability in deep learning models.It proposes a Multi-Grained Attention Recommendation(MGAR)model that integrates rating data with textual information(reviews,product descriptions).Utilizing two parallel Convolutional Neural Networks(CNNs)to process user and item reviews and descriptions,the model employs word-level,phrase-level,and sentence-level attention mechanisms to extract semantic information across various granularities.This facilitates a deep integration of review text,enabling a more effective capture of user preferences and product characteristics.Experimental results on four Amazon datasets indicate that the proposed model outperforms traditional rating-based recommendation systems and the recently popular DeepCoNN model in terms of predictive accuracy.

关 键 词：推荐算法 多粒度注意力机制 卷积神经网络 评论文本 数据稀疏性 

分 类 号：G63[文化科学—教育学]