Research On Single Domain And Cross Domain Recommendation Algorithm Based On Capsule Network

Many recommendation systems use reviews as auxiliary information to alleviate the problem of data sparsity and cold start.In the existing research,the depth models used to mine review features,such as convolutional neural network and attention,will lose some important information of the text,and the use of all reviews will also make the model affected by the low-quality reviews.In addition,the method of transferring the knowledge learned from other domains to the target domain also ignores some influence of multiple source domains on the target domain recommendation.Based on the above problems,this paper models from two angles of single domain recommendation and cross domain recommendation,and constructs two recommendation models based on capsule network.First of all,in view of the fact that the user has certain interaction in the target domain,this paper proposes self-attention capsule network rating prediction model with review-quality: SACR.SACR uses self-attention capsule network to mine the fine-grained features in the review document,uses the id number of users and items to mark low-quality reviews,and combines the representation of the two to predict the rate.The model also improves the squeeze function of capsules,so as to get more accurate high-level capsules.The model solves the problem of sparsity of rating data in the target domain and the problem of low quality reviews affecting the accuracy of recommendation.Secondly,in view of the cold start of users in the target domain,this paper proposes a cross domain rating prediction based on aspect and capsule network: ACN.ACN represents each user in the source domain as a review document and each item in the target domain as a review document.The model improves the mapping matrix of the capsule network and uses it to mine the user aspects and item aspects in the reviews,selects the most important features in the target domain through the attention mechanism,and transfer them to the target domain to predict the rate.This model can learn and transfer the knowledge of different source domains and multi-source domains,solve the problem of incomplete information mining under user cold start,and improve the accuracy of cross domain recommendation.Finally,the SACR and ACN proposed in this paper are compared with benchmark models in real-world datasets to verify the effectiveness and accuracy of the models.