Research And Application Of Recommendation Algorithm Based On Graph Neural Network

The recommender system can provide users with interesting items and play an important role in Internet applications.As an emerging graph representation learning method,graph neural network can generate low-dimensional feature representations for users and items based on graph structure,and then provide feature input including node neighbor structure information for recommender system.The existing graph neural network methods are widely used in recommender systems and have many progresses,but many of them still have the following problems: 1.The uniform sampling method for neighbor nodes may ignore the information provided by important neighbors;2.No suitable learnable parameter provided for rating prediction,which limits the application effect of graph neural network in the recommender system;3.Having not considered the timing characteristics of the neighbor structure of the graph in the recommender system;4.Ignoring the shallow output of the middle layer of the graph neural network.In response to the above problems,this thesis has completed the following works:1.Proposed a graph neural network model with rating-confidence-based sampling.The model includes: 1)A neighbor sampling method based on rating confidence,which takes into account both the time information of the rating and the degree of nodes in the graph structure,so that the sampling process can be biased towards those more valuable nodes.2)A rating prediction method based on the attention mechanism,which measures the importance of the various dimensions of the user and item characteristics in the scoring prediction task,so that the application effect of the graph neural network in the recommender system is improved.This thesis verifies the effectiveness of the model on six public data sets.Compared with several existing baseline models,the MSE value of this algorithm can decrease by 27.89%,and the MAE value can decrease by 26.01%.2.Proposed a gated graph neural network model with ordered input.This method can be effectively applied to the recommender system.The main contributions of this method include: 1)Pay attention to the timing characteristics of node neighbor structure.The neighbor nodes are input to the gated graph neural network in chronological order,and then the context capture ability is used to capture the timing characteristics in neighbor institutions.2)Pay attention to the shallow state output of the middle layer.Multi-head attention mechanism is used to fuse multi-layer state output information,so that the shallow structural features provided by the middle layer can better participate in the scoring prediction task.This thesis conducts experiments on six actual data sets.Compared with the optimal baseline,the MSE loss of the model on the four data sets can decrease by 13.36%,8.03%,27.49% and 2.52%,respectively.3.Finally,two network representation learning methods are applied to the recommendater system to realize book recommendation.Firstly,we analyzed and preprocessed the book data,build a rating-based recommendation system,generated feature vector representations for users and items based on the graph neural networks.Then we further implemented rating predictions,according to the ranking results of the predicted ratings.Finally,we recommended several items for each user.The front and back ends of the system are separated,and the web development framework is based on Django,VUE and Bootstrap.