| In this era of information overload,personalized recommendation systems are playing an increasingly important role as a bridge between users and massive amounts of data.However,as more stringent privacy policies are implemented,collecting user behavior data has become difficult,resulting in decreased effectiveness of traditional recommendation algorithms.In this context,session-based recommendation systems have attracted increasing attention from both academia and industry,aiming to predict the next item that may appear in an ongoing session without using any user profiles or historical data.This article proposes the Dynamic Embedding Network model and the Hyperbolic Hypergraph Convolutional Network model from two perspectives,namely the importance and transition relationships of items in session-based recommendation,respectively,to address some shortcomings of previous research.The main research contributions of this article are summarized as follows:(1)Regarding the importance of items,the article proposes the DEN model,based on contrastive learning and dynamic representation neural network,which captures the main preferences of users for session-based recommendation using an affinity matrix and effectively distinguishes unrelated items in sessions.The article also mitigates the model’s overfitting risk with a self-supervised contrastive learning auxiliary task.(2)Regarding the transition relationships of items,the article proposes the HHCN model,based on hyperbolic graph convolutional neural network.This model analyzes the inadequacy of previous research in capturing item transition relationships in sessions and uses hyperbolic graph convolutional neural networks to capture complex item transition relationships in sessions.By mapping the model and data to hyperbolic space,the article captures the correlation and hierarchical structure representation of user behavior in sessions more effectively,and alleviates the problem of data sparsity caused by the hypergraph structure,allowing the model to perform better in low-storage scenarios.(3)The article conducts extensive experiments on three real datasets,Tmall,Diginetica,and Nowplaying,and compares and analyzes them with common baseline models in recent years.The article also verifies the effectiveness of the two proposed innovations through ablation experiments on these datasets. |
