基于双注意力机制和迁移学习的跨领域推荐模型  

A Cross-Domain Recommendation Model Based on Dual Attention Mechanism and Transfer Learning

作　　者：柴玉梅[1] 员武莲 王黎明[1] 刘箴[2] CHAI Yu-Mei;YUN Wu-Lian;WANG Li-Ming;LIU Zhen(School of Information Engineering,Zhengzhou University,Zhengzhou 450001;School of Information Science and Technology,Ningbo University,Ningbo,Zhejiang 315211)

机构地区：[1]郑州大学信息工程学院,郑州450001 [2]宁波大学信息科学与工程学院,浙江宁波315211

出　　处：《计算机学报》2020年第10期1924-1942,共19页Chinese Journal of Computers

基　　金：国家自然科学基金(U1636111)资助.

摘　　要：跨领域推荐可用于解决单一领域数据稀疏导致的推荐系统性能退化问题,还可以缓解推荐系统中存在的用户冷启动问题.然而,现有的方法大多利用用户对项目的评分进行建模,忽略了评论文本所蕴含的信息.为此,本文提出了一种基于双注意力机制和迁移学习的跨领域推荐模型,首先通过CNN对评论文本建模,提取用户和项目特征;其次通过构造融合词的上下文关系的词注意力机制从评论文本中捕获词级别的信息,以提升CNN对文本中重点信息的关注度;然后通过构造特征突显机制从CNN提取到的用户特征和项目特征中捕获特征级别的信息;最后引入迁移学习,通过同时提取领域特有的特征和领域间的共享特征进行不同领域之间的联合建模,进行评分预测.本文在Amazon数据集上进行了实验比较与分析,首先对本文模型的推荐性能进行评估,与现有的跨领域推荐模型相比,在两种不同的跨领域数据集上平均绝对误差分别提升6.1%和9.15%,均方根误差分别提升3.66%和7.01%;然后对本文模型的知识迁移性能进行评估,与现有的单领域推荐模型相比,在不同数据集下均方误差分别提升5.47%和10.35%;最后通过实验验证了本文提出的注意力机制的有效性,及在缓解数据稀疏问题和用户冷启动问题方面的优势,也验证了模型的普适性.Cross-domain recommendation can be used to solve the problem of degrading the performance of the recommendation system caused by sparse data in a single domain,and it can also alleviate the cold start problem of users in the recommendation system.However,most of the existing methods use the user’s rating data to model the item,ignoring the review text written by the user for the item and the rich user and item information it contains.In recent years,deep learning has been successfully applied to various fields.Inspired by this,this article proposes a Cross-Domain Recommendation Model based on the Dual Attention Mechanism and Transfer Learning(AMTR)based on the review text.Firstly,modeling review text through convolutional neural network,extract user and item features.Secondly,this paper constructs the word attention mechanism and feature highlighting mechanism that fused the context of words,and proposes a feature extraction network based on the dual attention mechanism.The word attention mechanism captures word-level information from the review text to increase CNN’s attention to the key information in the text,while making the recommendation interpretable;The feature highlighting mechanism captures feature level information that is helpful for rating prediction from user features and item features extracted by CNN.Finally,based on the feature extraction network,transfer learning is introduced to jointly model between different domains to achieve knowledge transfer between domains.In this process,feature extraction networks in different domains are used to simultaneously extract domain-specific features and share features and combine with factorization machine to perform rating prediction to achieve cross-domain recommendation.In this paper,the experimental comparison and analysis is carried out on the Amazon dataset.The experimental results show that the performance of the AMTR model in terms of mean absolute error,root mean square error and mean square error is better than the comparison model.Firstly,eval

关 键 词：推荐系统 迁移学习 跨领域 CNN 评论文本 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]