Research On Recommendation Algorithms Based On Graph Neural Network

In recent years,recommender systems have been widely studied,among which,the research of graph neural network-based recommendation models is one of the hot spots.Many graph neural network-based recommendation models have been proposed in recent years.Although these models have achieved impressive performance,there are still some shortcomings.First,in the Top-N recommendation task,the existing graph neural network-based recommendation models fail to sufficiently exploit two important relationships,which makes the models fail to accurately model the preferences of the users and hinder the performance of existing graph neural network-based recommendation models.Second,in the session-based recommendation task,existing graph neural network-based session recommendation models do not make the messages propagating between items and auxiliary information,which makes the models cannot fully explore the relationship between the items and the auxiliary information,limiting the recommendation performance.In this thesis,we conduct an in-depth study and propose two recommendation models to address the shortcomings.Comparative experiments conducted on widely used real-world datasets confirm the effectiveness of the recommendation models proposed in this thesis.The main contributions of this thesis are as follows.1.The literature of Top-N recommendation models,session-based recommendation models,graph neural network-based recommendation models,etc.,is investigated.An in-depth analysis of existing graph neural network-based recommendation models is conducted.The shortcomings and possible methods of improving are pointed out.2.A dual-message propagation mechanism is proposed.The mechanism first constructs a dual-link user-item graph,and then defines preference messages and similarity messages for the preference relationship and similarity relationship,respectively.The special propagation mechanism can make both user-item preferences and similarity among users/items be represented explicitly,which is convenient for graph neural networks to model them directly.3.Based on the dual-message propagation mechanism proposed in this thesis,a recommendation model based on the dual message propagation mechanism,DGCF,is proposed.Unlike existing models that use only one graph neural network,DGCF uses two graph neural networks to model the user-item preferences and similarities among users/items.Compared with existing models,DGCF exploits these two types of relationships more deeply,achieving better recommendation results.Although DGCF uses two graph neural networks,its computational complexity does not increase significantly.4.A recommendation model based on item-category message propagation,IPGNN,is proposed.Unlike existing models that construct the item graph and category graph separately,IPGNN unifies the item sequences and category sequences to construct an item-category graph and then performs message propagation between item nodes and category nodes.IPGNN then models the relationships between items and categories using graph neural networks.Since the message propagation between items and categories is performed,IPGNN sufficiently exploits the relationships between items and categories and achieves better recommendation results.5.In this thesis,the two proposed recommendation models,DGCF and IPGNN,are compared with the existing advanced recommendation models on widely used real-world datasets.The experimental results verify the effectiveness of the recommendation models proposed in this thesis.