Research On Recommendation Algorithm Based On Graph Structure Network

With the explosive growth of Internet data,people are immersed in an ocean of information,they have to read and process large amounts of information mixed with redundant and irrelevant content in daily life.In the face of the increasingly serious problem of “information overload”,people always struggle to locate the information they require quickly and accurately.Recommender systems have been identified as an effective approach to mitigate the challenges described above.The function of these systems is to recommend content or items to users based on personalized information such as historical behavior and preferences,thereby enhancing user satisfaction and experience.Due to the fact that the interaction between users and items can be modeled as a bipartite graph,graph neural networks possess a stronger ability to extract information from graph-structured data,and have been widely used in the field of recommendation algorithm recently.Nevertheless,there still exist numerous challenges in recommendation algorithm based on graph structure networks.Firstly,recommender systems usually deal with a large amount of user and item data,where the interactions between users and items are frequently sparse,inevitably leading to data sparsity issues.Secondly,recommender systems typically rely on observable user interactions to make recommendations,which hinders the effective utilization of implicit relationships between users and items.Moreover,the sampling of original interaction data samples is often insufficient.Thirdly,existing recommender systems lack effective modeling techniques to capture high-order and complex relationships between users and items,resulting in incomplete user feature learning.In light of the aforementioned situation and challenges,addressing the limitations in existing recommendation algorithms based on graph-structured networks,the main research objectives of this paper are as follows:(1)Existing recommendation algorithms based on graph convolutional networks mainly focus on improving the model structure,ignoring the importance of improving the sampling quality of original samples and mining implicit relationships between users and items.To address this issue,this paper proposes a graph contrastive learning recommendation algorithm based on mixed sampling(MSGCL).Firstly,the algorithm uses a mixed sampling method to extract part of the information in positive samples and inject them into negative samples,thereby generating new informative hard negative samples.Secondly,to extract features on hard negative samples,the algorithm uses the light graph convolution network to obtain node representations of users and items.Moreover,the algorithm carries out a neighborhood contrastive learning method to mine implicit relationships between users and items.Finally,a multi-task learning strategy is adopted to jointly optimize the recommendation supervision task and contrastive learning task.The algorithm is evaluated on two real datasets Yelp2018 and Amazon-book,using Recall and NDCG evaluation indexes,and experimental results demonstrate its superiority.(2)Existing group recommendation algorithms based on hypergraph convolutional networks often overlook group similarity,make it difficult to model high-order complex relationships among users.Additionally,group recommendation suffers from data sparsity issues due to sparse group interactions.To address these issues,this paper proposes a group recommendation algorithm based on hypergraph contrast learning(HCLGR).Firstly,a userlayer hypergraph is defined to extract high-order relationships of users and aggregate user preference features across three different channels.Secondly,a group-layer hypergraph is defined to connect all groups into an overlapping network,further learning group preferences.Finally,a training strategy combining collaborative training and contrastive learning is adopted to optimize the model and achieve data augmentation.The model is evaluated on real datasets Douban and Weeplaces,using Recall and NDCG evaluation indexes,and experimental results verify the effectiveness of the proposed model.