Recommendation Algorithms Based On Graph Neural Networks

Personalized recommendation system is an important means to solve information overload.Through the analysis and mining of user historical behavior logs,user personal information,and item attribute information,it learns users' interests and preferences and actively recommends items that may be of interest to users.The recommendation algorithm is the core of the personalized recommendation system and determines the performance of the recommendation system to a large extent.Traditional recommendation algorithms generally act on structured data,and the feature extraction from unstructured data is not sufficient,which limits the accuracy of recommendation results.There are a variety of unstructured data with graph structure in the recommendation system,and the graph neural network shows a powerful feature extraction ability on the data with graph structure,which provides a new idea for the research of recommendation algorithm.This paper studies the recommendation algorithm based on graph neural networks,and designs effective graph neural networks,which are applied in a variety of recommendation scenarios to improve recommendation performance.The specific work is as follows:(1)An item rating prediction method based on combined multi-receptive field graph convolutional network is proposed.The method as a whole has an end-to-end structure.First,the user-item rating matrix is transformed into a bipartite user-item graph with weight.Then,through multiple graph convolutional layers with different receptive fields,the multi-level embedding representations of user nodes and item nodes are obtained,and the final embedding representations of users and items are obtained by combining them.Finally,the final embedding of the user and the item is used to obtain the user's rating of the item through the bilinear decoder.The experimental results show that this method alleviates the oversmoothing phenomenon that easily occurs in graph convolutional neural networks to a certain extent.Compared with a variety of existing item rating prediction models,this method can more fully capture the structural features between users and items,and obtain more accurate user embedding representations and item embedding representations,thereby improving the accuracy of scoring predictions.(2)A multi-level influence diffusion Network for social recommendation is proposed.This method obtains the initial embedding representation of the user and the final embedding representation of the item through the fusion layer.A multi-level influence diffusion module is proposedthat contains several branches composed of multiple graph convolutional neural networks,which is used to simulate the diffusion process of users' social influence in social networks.The user's social embedding representation is obtained through the attention mechanism.At the same time,the user's interest embedding expression is obtained from the user's historical preferences.Then the user's social embedding representation and interest embedding representation are summed up with corresponding elements to get the user's final embedding representation.Finally,a recommendation list is generated for the user through the final embedding representation of the user and the final embedding representation of the item.Experimental results show that this method can more accurately simulate the diffusion process of influence in social networks,and obtain more accurate user embedding representation,improving the accuracy of recommendation.Compared with a variety of existing social recommendation methods,this method can generate a more accurate recommendation list for users.(3)A multiple gated graph neural networks with different timesteps for session-based recommendation is proposed.This method first constructs the session sequence as a session graph.Then,multiple gated graph neural networks with different time steps are used to capture the complex transfer relationship between items,and the multi-level embedding features of item nodes are obtained.The long short-term memory network is used to aggregate the multi-level embedding features of item nodes.Finally,the user's general interest preferences are obtained through the attention mechanism,and combined with the user's current interest preferences to obtain a session-level embedding representation,thereby generating a recommendation list for the session.Experimental results show that this method can more accurately model the complex transfer relationship between items.Adopting long short-term memory network aggregation method effectively alleviates the over-fitting problem in graph neural network,and obtains more accurate item node embedding representation and session embedding representation.Compared with a variety of existing session-based recommendation methods,this method can generate a more accurate recommendation list.