Collaborative Recommendation Based On Multi-graph Fusion

Personalized recommendation is dedicated to quickly discovering and accurately capturing user preferences,improving user experience,and increasing user stickiness to the platform.Collaborative filtering is one of the most important recommendation techniques and plays a pivotal role in personalized recommendation.Collaborative recommendation,which relies only on user-item interaction,greatly limits recommendation performance due to data sparsity.Most data in recommendation scenarios naturally have graph structures,such as graphs of user behavior sequences,graphs of social relationships among users,and knowledge graphs among item attribute information.The fusion of these multi-graph information can effectively compensate the shortage of recommendation based on user-item interaction information,provide additional feature information to the recommendation system,and improve the accuracy of recommendation.In this thesis,with the powerful propagation and expression capabilities of graph neural networks,two different multigraph fusion collaborative recommendation algorithms are proposed for general recommendation scenarios and temporal recommendation scenarios,respectively,as follows:(1)For the general recommendation scenario,most existing graph-based recommendation algorithms rely only on user-item heterogeneous graphs for collaborative recommendation.In fact,user-user and item-item homogeneous graphs also contain information that is very important for recommendation,which can compensate for the very sparse user-item bipartite graph.Therefore,this thesis proposes a multi-graph collaborative recommendation framework with iterative fusion of heterogeneous and homogeneous graphs to improve the embedding quality by iteratively fusing user-item heterogeneous graphs,user-user homogeneous graphs,and item-item homogeneous graphs.The multi-information intersection module and the multi-graph fusion propagation module are the core of the framework.The former achieves dual feature intersection among heterogeneous nodes throughout the embedding process;the latter iteratively integrates homogeneous nodes(users or items)and their topological relationships based on the constructed user-user and item-item homogeneous graphs.Extensive experiments on real datasets show that the proposed recommendation model outperforms other algorithms and models,and a series of ablation and parameter analysis experiments also verify the effectiveness of the components in the framework to iteratively fuse multiple information and improve the node embedding representation accuracy.(2)For the temporal recommendation scenario,most existing temporal recommendation algorithm approaches focus on user-item interactions and ignore the widespread sequential patterns in user behaviors,resulting in the inability to fully utilize historical interaction information to obtain dynamic user and item representations.Dynamic behaviors and historical interactions are crucial to achieve temporal recommendation.Therefore,this thesis proposes a multi-graph temporal recommendation framework with the fusion of interaction and sequence graphs.To better capture fine-grained temporal behavior,the framework reconstructs loose sequences of items in user behavior into sequence graphs of item-and item-directed interactions through metric learning.The sequence graph is then fused with the global interaction graph to extract the actual interests of users from the noisy behavioral sequences and explore the user-item interaction behavior.Extensive experiments on real datasets show that the proposed recommendation framework outperforms other baseline recommendation algorithms and models,and a series of ablation and parametric analyses also show that the model components can better capture finegrained behavioral information of users,dynamically model user preferences,and fuse interaction graphs and sequence graphs to improve the embedding quality.The thesis has a total of 14 figures,11 tables,and 105 references.