Research On Hybrid Recommendation Model Based On Knowledge Graph Embedding

In order to solve the problem of information explosion in various online applications and improve user experience,a recommendation system is proposed to model user preferences.Despite efforts to make recommendations more personalized,recommender systems still face challenges such as sparse data and cold start.In recent years,knowledge graph(KG)as an auxiliary information generation recommendation has aroused great interest.This approach not only alleviates the above problems and provides more accurate recommendations,but also explains the recommended items.The main work and innovation in this thesis are reflected in the following three aspects:1.In view of the fact that the current KG recommendation model does not pay enough attention to the message propagation interaction in the process of transmission in the form of multi-KG,and ignores the characteristics of the change of the user’s latent intention,in this thesis,a Collaborative Interaction Knowledge-aware Network(CIKN)recommendation model is proposed.The model uses auxiliary KG to explore the propagation interaction strategy and explore the potential user intention behind the user-item interaction.This article designs a a copropagation interaction layer consisting of item-side and user-side KG to explicitly encode the collaborative signal.At this layer,the user’s intention is constructed by constructing the userside KG;then,the intent representation is extracted into the item-side KG to learn the user and item representation together.2.On the basis of enriching the representation of users and items,this thesis proposes a new Deep Field-Aware Interaction Machine(DeepFIM)recommendation model,which solves the problem of "short-expression" of features in the feature field and better captures multiintensive feature interaction.This thesis designs an interaction layer to identify intrafield and interfield feature interactions,and use attention mechanism to distinguish the importance of different features.In particular,this thesis employs two embedding modes for field-aware embedding queries;this will be more conducive to learning interfield and intrafield characteristics.In addition,this thesis also introduces dynamic Bi pool layer to enhance the acquisition of high-order features,so as to maximize the retention of information,which is conducive to the subsequent learning of deep neural networks.3.Using the open datasets of music,books,movies and restaurant recommendations,the research theory based on the knowledge graph embedding recommendation model proposed in this thesis is tested and analyzed.The experimental results show that the proposed model is more effective and reliable than the existing methods in the recommendation performance.