Research On News Recommendation Model Based On Hypergraph

In the Internet era,numerous news articles are generated every day,and personalized news recommendation for users is a very popular research direction.Personalized news recommendation recommends news articles based on the user’s interests by learning from the news and user representations.The feature extraction of news and users is a critical problem in personalized news recommendation.Using graph neural networks to extract features of users and news is a mainstream research direction,and many graph neural network-based news recommendation models have achieved good results in feature extraction.However,there are still some problems.This thesis proposes a Hypergraph Transformer model for feature extraction of news and users.The model optimizes the attention structure of Transformer using hypergraph neural networks to improve representation learning of users and news.Self-supervised learning is used to enhance the model,and the probability of users clicking on news is ultimately obtained,and news is ranked using ranking learning.The main contributions of this thesis are as follows:1.To address the problem of data sparsity and skewness in current news recommendation,this thesis proposes a Hypergraph Transformer-based news recommendation model that optimizes the modeling of news and users.The model utilizes the powerful representation learning capabilities of hypergraphs to mine user and news features deeply.A local graph structure is constructed using graph neural networks,and adaptive hypergraph relationship learning is used to enhance user collaboration modeling to alleviate data sparsity issues.Transformer is used to transfer information from dense user or news nodes to sparse user or news nodes to solve the problem of skewed data distribution.Finally,the propagation and backpropagation between nodes and hyperedges are used to update user and news embeddings,and the model shows good performance in the experimental dataset.2.To address the problem of noise in interactive data that affects the graph topologyaware embedding of local collaborative modeling,this thesis uses self-supervised learning to optimize the news ranking module on the basis of the Hypergraph Transformer model,reducing the impact of noise in interactive data.To solve this problem,selfenhancing learning is used to enhance model training between local topology-aware embedding and global hypergraph learning.Information is transferred from the highorder denoising features in Hypergraph Transformer to the low-order noisy topologyaware embedding,the local graph structure is recalibrated,and the model’s robustness is improved.By using meta-learning to construct reliability labels,the embedding is scored for reliability,and the reliability score is applied to the news ranking,which shows good performance in the experimental dataset.