基于双塔隐语义与自注意力的跨域推荐模型

Cross-domain recommendation model based on two-tower latent semantic and self-attention

为缓解跨域推荐数据稀疏与冷启动问题,该文提出一种融和双塔隐语义与自注意力机制的跨域推荐模型(DLLFM-DA/Self atten CDR model,DLDASA),能够有效提升目标域推荐准确率.首先利用提出的双塔隐语义模型(DLLFM),借助源域和目标域用户的类别偏好和项目类别,生成高质量隐语义;其次,在隐语义特征迁移过程中引入域适应(domain adaptation),有效对齐源域与目标域的特征分布,最小化源域与目标域间数据分布差异,提供更高质量的隐语义特征迁移;然后利用多头自注意力机制捕捉两个域之间的差异性与相关性,对差异信息进行筛选与融合,缓解负迁移问题,以提升跨域推荐质量;最后,在Movielens-Netflix和一品威客(YPWK)-猪八戒网(ZBJW)真实数据集上,将DLDASA与基线单域和跨域推荐模型进行对比实验,结果表明,均方根误差(RMSE)和平均绝对误差(MAE)均有明显改善.该研究验证了DLDASA模型能够更充分地提取用户特征,有效缓解目标域信息不足的问题.

To alleviate the sparse data and cold start problem of cross-domain recommendation,a cross-domain recommendation model(DLLFM-DA/SelfAtten CDR Model(DLDASA))is proposed to integrate the dual-tower hidden semantics and self-attention mechanism,which can effectively improve the target domain recommendation accuracy.First,it uses the proposed dual-tower hidden semantic model(DLLFM)to generate high-quality hidden semantics with the help of category preferences and item categories of users in the source and target domains;second,it introduces domain adaptation in the process of hidden semantic feature migration to effectively align the feature distribution between the source and target domains,minimize the difference of data distribution between the source and target domains,and provide higher quality hidden semantic feature migration.Then,the multi-headed self-attentiveness mechanism is used to capture the difference and correlation between two domains,and filter and fuse the difference information to alleviate the negative migration problem in order to improve the quality of cross-domain recommendation.Finally,the experimental results of comparing DLDASA with the baseline single-domain and cross-domain recommendation models on Movielens-Netflix and YPWK(YPWK)-ZBJW(ZBJW)real datasets show that both root mean square error(RMSE)and mean absolute error(MAE)are significantly improved.It is verified that the DLDASA model can more fully extract user features and effectively alleviate the problem of insufficient information in the target domain in this study.