Personalized Recommendation Algorithm Combining Deep Learning And Knowledge Graph

With the explosive development of the internet,the information load in the network has increased significantly,leading to the emergence of recommendation systems.Despite the rapid increase in the overall data volume and types of information on the network,various types of user interaction information still suffer from insufficient data,resulting in personalized recommendation results that are not satisfactory.To compensate for the lack of interaction information data,researchers often introduce auxiliary information to improve recommendation performance.Among various types of auxiliary information,knowledge graphs are a widely used heterogeneous graph structure information in recommendation systems,with rich semantic and structural features.However,current knowledge graph-based recommendation systems face several challenges,including:(1)The input end lacks sufficient user-item interaction data,resulting in a very sparse user-item interaction graph that affects the accuracy of recommendation results;(2)Knowledge graph-based recommendation systems usually consider the relationship features between entities in modeling,but ignore the entity features themselves,leading to low utilization of feature information.To address these challenges,there are several commonly used methods,including:(1)Combining historical user behavior data to better explore hidden user interests;(2)Using the subsumption and attribute relationships in the knowledge graph to further enrich the feature representation of items.This paper further integrates deep learning techniques into these two commonly used methods,while improving the model’s generalization and robustness.This paper aims to optimize the construction of different knowledge graph network structures using neural network technology in deep learning.The introduction of neural network technology retains the advantages of knowledge graphs while utilizing the connection relationships between entities.By exploring the neighborhood relationships of user-interacted items,the model can depict more accurate user profiles and provide more suitable services for users.Two different recommendation algorithms are proposed in this paper,with the following main contributions:(1)This paper proposes a novel knowledge graph recommendation algorithm,named Message Passing and Preference Propagation Knowledge graph Recommendation(MPPPKR).The model first constructs an initial knowledge graph based on the user’s historical interaction data.Then,two types of neighborhood topologies are modeled on the initial knowledge graph,and an alternating relationship message passing algorithm is utilized to integrate them together for predicting unknown relationships.With the predicted unknown relationships,the model completes the knowledge graph and builds a preference propagation network on it to explore users’ hidden interests.The preference propagation algorithm is employed to extract the multiorder response embedding of user’s click history for a given item,and then combine them to obtain the user feature embedding.Finally,the model predicts the click probability of a given item by utilizing both the user and item feature embeddings.Experiments on real-world movie and fashion e-commerce datasets demonstrate that the proposed model outperforms several baseline models,including CKE,CFKG,and MKR,in terms of prediction accuracy,which verifies the effectiveness of the proposed model.(2)This paper proposes a novel knowledge graph recommendation algorithm called Message passing and Attention-based Graph convolution Knowledge graph Recommendation(MAGKR).First,the model completes the knowledge graph in a consistent way.Then,based on the completed knowledge graph,an attention-based graph convolutional network is constructed.The network uses layer-wise and user-wise representation learning methods to obtain the final user feature embeddings.The convolutional block attention module is adopted to refine the feature tensor representation in the representation learning method.The model considers both entity interactions and layer interactions when dealing with various types of feature information and aggregates features of different forms using attention mechanisms.Finally,the model predicts the interaction probability between users and given items based on the user feature embeddings and item feature embeddings.Experimental results on real datasets in the domains of audio-visual and fashion e-commerce show that the proposed model has certain advantages over current state-of-the-art models such as AFM,NCF,and DCNV2 in terms of accuracy in project recommendation prediction.This paper proposes two improved knowledge graph recommendation algorithms that incorporate deep learning techniques.Through experimental comparison,both algorithms outperform the current baseline model in click-through rate prediction recommendation,indicating that the two models can effectively suppress the negative impact caused by sparsity and cold start problems,and can efficiently utilize various entity features to enhance the personalized recommendation accuracy for users.The results suggest the potential of utilizing deep learning in knowledge graph recommendation systems.