摘要
为缓解跨域推荐中目标域数据稀疏和冷启动问题,综合增强嵌入、嵌入迁移、注意力机制调整和跨域推荐技术,提出一种融合深度特征提取和注意力机制的跨域推荐模型(cross-domain recommendation model of deep feature extraction and attention mechanism,CRDFEAM).利用潜在因子模型将类型相似度合并到矩阵分解过程,挖掘项目类型的隐性偏好.相比评分这一显性偏好,项目类型能更充分获取用户特征.在跨域迁移时,用分布对齐方式使域间差异最小化,以减少两个领域特征之间的数据分布差异.相对于直接迁移,分布对齐方式具有更强的可解释性.在特征调整过程中,引入多层感知机(multilayer perceptron,MLP)映射,并使用注意力机制进一步调整用户特征,使源域中没有出现过的目标域用户注意到源域用户的特征信息,同时也使源域中出现过的目标域用户注意到目标域中的项目特征信息.在真实数据集Movielens(M)、Netflix(N)和Douban(D)上的实验验证结果表明,引入MLP映射嵌入的CRDFEAM+模型的均方根误差(root mean square error,RMSE)值较基准模型跨域潜在特征映射(cross-domain latent feature mapping,CDLFM)平均提升9.88%,平均绝对误差(mean absolute error,MAE)值平均提升11.14%.研究验证了CRDFEAM+模型的跨域推荐效果,能够更充分地提取用户特征,有效缓解目标域信息不足问题.
To alleviate the problem of sparse target domain data and cold start in cross domain recommendation,Cross-domain recommendation model of deep feature extraction and attention mechanism(CRDFEAM)model is proposed by combining the techniques including the enhanced embedding,embedding transfer,attention mechanism adjustment,and cross-domain recommendation.Firstly,the type of similarity is merged into matrix decomposition process to mine the implicit preference of project type.In contrast to the explicit preference of user rating,the implicit preference can obtain the user characteristics more completely.Secondly,to reduce the difference in data distribution between different domain features during data migration,the distribution alignment method is used to minimize the difference between the domains,which is more interpretable than the direct migration.Finally,for the feature adjustment,the multilayer perceptron(MLP)mapping is invoked and the attention machine is used to adjust user features.In this way,the target domain users who have not appeared in the source domain can notice the feature information of source domain users,and the target domain users who have appeared in the source domain can also notice the feature information of target domain items.The purpose of above mentioned operations is to alleviate the sparse problem of data in the target domain.In summary,compared with the benchmark cross-domain latent feature mapping(CDLMF)model,the root mean square error(RMSE)value of CRDFEAM+increases by 9.88%on average,and the mean absolute error(MAE)value increases by 11.14%on average.The research verifies the good cross-domain recommendation effect of the CRDFEAM+model,which can more fully extract user features and effectively alleviate the problem of insufficient target domain information.
作者
操凤萍
张锐汀
CAO Fengping;ZHANG Ruiting(Department of Computer Engineering,Southeast University Chengxian College,Nanjing 210088,Jiangsu Province,P.R.China;Nanjing Jiangdong Middle Road Huatai Securities,Nanjing 210019,Jiangsu Province,P.R.China)
出处
《深圳大学学报(理工版)》
CAS
CSCD
北大核心
2023年第3期266-274,共9页
Journal of Shenzhen University(Science and Engineering)
基金
国家重点研发计划资助项目(2020YFC2007401)
江苏省重点实验室资助项目(2242021K30021)。
关键词
人工智能
迁移学习
跨域推荐
注意力机制
特征嵌入
潜在因子
矩阵分解
artificial intelligence
transfer learning
cross-domain recommendation
attention mechanism
feature embedding
potential factors
matrix factorization