| In the present times,with the fast growth of e-commerce,social media and other domains,recommendation system has become one of the most popular and important applications.Personalized recommendations aim to provide users with suitable items based on their interests and preferences,and the technologies associated with recommendations have garnered extensive interests from academia and industry.As a paradigm for building recommendation models,collaborative filtering can learn users’ interests and preferences from the collected historical users behavior data to make efficient recommendations.However,most of the existing collaborative filtering recommendation methods only focus on a single type of interactive information,ignoring the diversity of user behaviors,which makes it difficult to capture the heterogeneous relationships between users and items across different types of interactive behavior,and also do not take into account the high-order connectivity between users and items,which is crucial for coding collaborative signals.In recent years,graph neural networks have gained significant attention owing to their excellent processing capabilities on graph structured data.The vast majority of data in recommendation system can essentially be regarded as graph structure data.Therefore,this thesis combines graph neural network technologies with multi-behavior recommendation task to propose two multi-behavior recommendation models based on graph neural networks,which are used to anticipating interests and preferences of users for items and augment the precision of recommendations.The research content and main contributions of this thesis are as follows:(1)As most existing recommendation methods only focus on a single type of interactive information between users and items,this thesis constructs a heterogeneous graph to represent multi-behavior data between users and items,and proposes a Multi-Behavior Weighted Graph Convolutional Network(MBWGCN)recommendation model.Firstly,the Graph Convolutional Network(GCN)embedding propagation layer with multi-behavior perception is designed,which jointly embeds users,items and edge relationships,and uses the high-order connections between nodes in the graph to capture potential collaborative signals in multi-behavioral interactions.Considering the different degrees of influence of different behaviors on users,propagation weight is set up to distinguish the importance of different behaviors.Secondly,a item correlation extraction module is set up to model the correlations between items,capturing multi-behavior semantics,and enhancing the learning ability for item embeddings.Finally,the embedding representations of nodes and relationships,as well as the correlation score between interactive items,are combined to obtain the final prediction score.This model has undergone extensive experiments on two publicly available datasets,and the experimental outcomes fully validate the effectiveness of the MBWGCN model.(2)Considering that interests of users are fundamentally dynamic and influenced by their historical sequence behaviors,multiple behaviors often exhibit time dependence.In order to more accurately capture users’ preferences,a Cross-Structure Multi-Behavior Contrastive Learning for Recommendation(C-MBR)model is proposed in collaboration with MBWGCN.The model considers both the graph-based global structure and the sequence-based local structure information,and multiple structures complement each other to capture the user’s interest preferences from different aspects.Specifically,under different behavior subgraphs,the model uses the corresponding encoder to obtain the user’s and item’s single-behavior embedding representations under the global structure and local structure,and then builds a multi-behavior fusion module to fuse the single-behavior embedding representations under different structures to obtain the final representations of users and items under the global and local structures.In order to further improve the accuracy of recommendation,C-MBR also includes two contrastive learning tasks,which model the commonalities between different behaviors and structures(local/global),in order to improve the quality of embedding representations and alleviate the problem of data sparsity.Finally,the model is jointly optimized by recommendation task loss function and contrastive loss function.The experimental outcomes show that the C-MBR model surpasses all baseline methods on two publicly available datasets and achieves the best performance,verifying the superiority of the C-MBR model. |
