Research On Personalized Recommendation Method Of Academic Papers Based On Deep Learning

In the era of the rapid development of the Internet,the publication rate of academic papers has increased dramatically,and the problem of paper information overload has become more severe for users than ever.Therefore,personalized recommendation of academic papers has become an effective method to solve this problem.Although a lot of scholars are working on how to improve the efficiency of personalized recommendation,the current recommendation systems are still facing a large number of problems.For example,the sparse data easily cause overfitting,the absence of a large amount of user data affects the accuracy of the recommendation results,and how to diversify the recommendation results.In order to address the above problems,this thesis proposes a deep learning-based recommendation model for papers,and the main research contents are as follows.First,this dissertation proposes a Self-attention and Residual mechanism hybrid model based on Collaborative Knowledge Graph(CKGSR).The model uses the Bi-LSTM network to get a comprehensive paper representation.In order to get deeper user reading preferences,the residual self-attention mechanism method is used to weight the user's neighborhood nodes.The bi-interaction aggregation method is used to fuse the user's neighborhood feature representation.Besides,the correlation between the user and the paper is analyzed and calculated using a multi-layer perceptron.The corresponding recommendation is finally derived results.This approach can measure the user's preference for different papers,thus personalizing the recommendation to the user and enhancing the interpretability of the recommendation.Second,the graph convolutional neural network is applied to CKGSR.A Graph Convolutional Network model based on CKGSR(CKGSRN)is proposed to realize the weighted higher-order propagation and aggregation of nodes and relations in CKG,and finally form a new node vector,which contains the neighborhood knowledge information in the graph.The new nodes include potential preferences of users,potential attributes of papers,etc.,thus helping to improve the diversity of recommendations while ensuring that comprehensive user and paper characteristics are obtained.Finally,based on the above model,a Deep Q Network model based on Collaborative Knowledge Graph and Graph Convolutional Neural Network(CKGND)is proposed to find the optimal recommendation strategy.The model utilizes the initial value module,which is reinforced to train and then initialize the recommender system to reduce the impact of early cold start on users;a deep Q-network is introduced,and the modules are trained to make a policy-optimal solution.This approach can effectively improve the user experience in the recommender system.Through comparative experiments on Cite ULike-a and Cite ULike-t datasets,it is found that the model proposed in this dissertation can perform better than classical recommendation models,which are based on collaborative filtering,knowledge graph,and reinforcement learning in terms of HR,NDCG,and MRR metric values.