Research On Graph Neural Network-Based Self-Supervised Learning Recommendation Algoriyhnm

Recommendation algorithms based on graph neural networks can better capture the association information between users and items by utilizing the user-item interaction graph,thereby effectively improving the performance of recommendation algorithms.However,there are still some shortcomings in existing works.On the one hand,the data sparsity problem is still the primary reason limiting the current algorithms to achieve higher recommendation accuracy.The current recommendation algorithms ignore the mining and utilization of auxiliary information and effective information in the unmarked interaction space.On the other hand,there is usually a popularity bias in the historical interaction data of users,and the current recommendation algorithms tend to recommend popular items by using imbalanced interaction data for training.Therefore,this paper proposes to use selfsupervised learning to improve the existing graph neural network-based recommendation algorithm,aiming to mitigate the impact of data sparsity and popularity bias on recommendation algorithms.Firstly,to address the problem of data sparsity in user-item interaction data that affects recommendation accuracy,the thesis propose an Item Knowledge-Aware Graph SelfSupervised Learning for Recommendation(IKGS)algorithm.This algorithm introduces the semantic relationships between items in the knowledge graph through a semantic-based item similarity graph and uses these semantic relationships to predict self-supervised signals in the unmarked interaction space,thereby alleviating the problem of insufficient interaction data.The algorithm improves the learning ability of item embedding representation by maximizing the mutual information between items in the interaction graph and selfsupervised signals.Finally,the model is further optimized through joint training of selfsupervised learning tasks and recommendation tasks.Secondly,in response to the problem of popularity bias affecting recommendation diversity,the thesis propose an Item Knowledge-Aware and Popularity Debiasing Graph Self-Supervised Learning for Recommendation(IKPGS)algorithm.This algorithm introduces a penalty constraint on item popularity during data augmentation of the user-item interaction graph,removing the popularity bias in the original interaction data.The algorithm avoids the influence of popularity bias when predicting self-supervised signals by constructing self-supervised learning tasks in multiple views.Finally,the algorithm improves the recommendation task loss function from the perspectives of both popular and non-popular items to enhance recommendation diversity through a multi-task learning strategy.Finally,the thesis verify the effectiveness of the two algorithms in mitigating data sparsity and popularity bias on three real datasets from various perspectives,including comparative experiments,ablation experiments of key modules,and parameter studies.