Research On Relation-aware Dual-tower Recommendation Model Based On Knowledge Graph

Knowledge graph(KG)as a kind of auxiliary information can effectively alleviate the interactive data sparsity and cold start problems of traditional recommendation algorithms.However,due to the heterogeneity and complexity of KG,a lot of noise will inevitably be introduced in the process of iteratively propagating user preferences indiscriminately over the entire KG.Besides,the recommendation algorithm based on knowledge graph lacks the consideration of fine-grained features of items under different relationships,and only focuses on how to mine item-side information while ignoring the extraction of user potential features,so the improvement of recommendation performance is not significant.In view of the above problems,this paper focuses on how to make full use of knowledge graph information and effectively combine it with the recommendation system to enhance the recommendation performance.The main work is as follows.Firstly,in view of the introduction of noise and the problem of model instability caused by the random sampling strategy,this paper proposes a relationship-aware sampling strategy,and selects a more representative sampling strategy through a more robust second-order adjacent importance sampling method.Neighborhood nodes to enhance the effectiveness of the model.Secondly,in order to effectively aggregate the selected neighborhood graphs under the different relation spaces,and then learn the latent feature of the target item in a specific relation space,a feature extraction algorithm RAAN based on relation-aware attention in knowledge graph is proposed.Relational factors are fused to differentially aggregate fine-grained features of items under different relationships to obtain more accurate semantic representations of items.The fine-grained features of item under different relationships are differentially aggregated by integrating relationship factors to obtain more accurate semantic representation of item.Thirdly,because the recommendation model based on the item-side knowledge graph lacks the mining of user related information,a method of fusing multi-feature information is proposed to model the user’s interest and preference features.And combined with the RAAN algorithm proposed in this paper,a multi-feature fusion dual-tower recommendation model is designed.On the basis of using knowledge graph as auxiliary information,the idea of collaborative filtering is integrated to model the information on both sides of user and item,so as to make full use of interactive data and knowledge graph to improve recommendation performance.Finally,the effectiveness experiments of the feature extraction algorithm fused with relation-aware attention and the dual-tower recommendation algorithm based on multi-feature fusion are carried out on three public datasets,and compared and evaluated in detail with various baseline models.