Research And Implementation Of Knowledge Graph Recommendation Algorithm Based On Contrastive Learning

As a technology that can provide personalized recommendations to users based on their interests and preferences,recommendation systems have played an important role in many online services such as e-commerce,video sharing,and online advertising,effectively alleviating the problems of difficult user selection and inefficient resource utilization caused by information overload.Among various recommendation techniques,collaborative filtering framework has become an effective solution for predicting user preferences.In recent years,collaborative filtering techniques have evolved from matrix-based decomposition to neural network-based,such as potential embedding learning using self-encoders and collaborative filtering using attention mechanism,etc.However,these methods still face the data sparsity problem when modeling user-item interaction patterns,i.e.,they cannot effectively deal with those users who have little or no interaction records with the project.To overcome the data sparsity problem,the knowledge graph is incorporated into the recommendation system as an external data that can demonstrate a large amount of semantic knowledge and rich contextual content to enhance the representation of users and items by encoding the semantic correlation between items and as useful auxiliary information to improve the recommendation quality.However,the success of knowledge graph recommendation methods relies on high-quality knowledge graphs and is susceptible to noise,which may not accurately learn high-quality representations,and still suffers from two shortcomings in practical applications: i)the longtailed distribution of entities leads to sparse supervisory signals for item representations enhanced by knowledge graphs;ii)real-world knowledge graphs usually contain noise and often contain between items and entities connections that are not related to the topic.Such knowledge graph sparsity and noise problems prevent item-entity dependencies from accurately reflecting their true characteristics,significantly amplify the noise effect,and hinder the accurate representation of user preferences.To improve these two problems,this paper designs a knowledge graph recommendation model based on contrast learning and applies the improved model to the Douban movie dataset to complete the system implementation.The main work of the full paper is summarized as follows:(1)In this thesis,A knowledge graph-based attention network recommendation model(KGAT)is used to address three major challenges of traditional collaborative filtering methods:ignoring other information about items,ignoring relationships between instances or items,and requiring domain-specific knowledge and manpower to construct paths.In this paper,we embed collaborative knowledge graphs into a low-dimensional space and capture higher-order relationships using an attention mechanism,while using multiple attention-embedded propagation layers to explore explicit higher-order connectivity and stitch higher-order relationships directly into a predictive model.Experimental results on both datasets show that KGAT outperforms current recommendation models in terms of accuracy,diversity,and interpretability.(2)In this thesis,A knowledge graph recommendation model(KGCL)based on contrastive learning is proposed to address two challenges faced in existing research: improving the recommendation effectiveness of long-tail items and enhancing the robustness of the model to noise in interaction data.This model introduces contrastive learning to collaborative knowledge graphs to enhance the representational and generalization capabilities of graph neural networks in knowledge graph recommendation systems,reinforces node representational learning through a self-discrimination mechanism,and jointly optimizes traditional supervised recommendation tasks and assisted contrastive learning tasks through multi-task learning strategy.Through extensive comparison experiments on three real datasets,it is demonstrated that KGCL can significantly enhance the performance,generalization ability and noise resistance of graph neural network-based knowledge graph recommendation.(3)Based on the model proposed in this thesis,a prototype knowledge graph recommendation system based on comparison learning is designed and implemented.The system takes the movie field as the application scenario,the underlying data is stored by My SQL and Neo4 J,and adopts a front-and back-end separation design pattern,providing a variety of recommendation strategies and cross-comparison functions.Through testing,it shows that the system has excellent performance in terms of recommendation accuracy,diversity and interpretability,which lays a good foundation for the subsequent development of a complete knowledge graph recommendation system based on contrastive learning.In summary,the knowledge graph recommendation model based on contrastive learning proposed in this paper provides an effective solution to solve the problems of data sparsity,long-tail effect and noise interference,combines contrastive learning with knowledge graph,and provides new ideas and methods to improve the effectiveness of recommendation systems.Meanwhile,the recommendation prototype system designed for the movie field also provides a useful exploration and attempt for the practical application of knowledge graph in recommendation systems.