Research On Collaborative Recommendation Model Of Academic Paper With Deep Feature

The rapid development of academic research and digital libraries has led to a sharp increase in academic papers on the Internet.How to provide personalized paper recommendation services for scholars has become a research hotspot in the field of data mining.Based on deep learning technology,this thesis conduct a research from enhancing paper and user features,and improve the recommendation performance of academic papers by integrating the deep features of papers and users in the collaborative filtering recommendation algorithm.The specific studies are as follows:(1)Based on the deep learning network of extracting common text features and the characteristics of the academic paper,this thesis designs and constructs a deep feature extraction network of paper content based on hierarchical attention mechanism.Because the abstract is streamlined and complete,it is used as the main input data of the network model,then the BERT and Recurrent Neural Network are respectively utilized to obtain the sentence features and paragraph features.Based on the key information of the paper carried by keywords and title,the attention model of wordlevel and sentence-level is constructed respectively,and the deep features in the abstract are extracted to accurately model the content of the paper,which lays a foundation for the recommendation of academic papers.(2)To solve the problem that rating-sparsity in collaborative filtering recommendation leads to inaccurate learning of paper features,this thesis proposes an academic paper recommendation model HAN-PMF that collaboratively learn the paper content and user ratings.Meanwhile,a corresponding multitask learning model is also proposed.Firstly,the deep feature extraction network based on the hierarchical attention of the thesis is used to enhance the paper latent feature vectors in PMF,so that it can learn the features of the paper from the user ratings matrix and the content of the paper.Then,the paper tag prediction task is added to the academic paper recommendation task,and the paper representation is further optimized by using the tag information.Related experiments have verified that the recommendation performance of this model is better than the current recommendation model of the same kind.(3)Rating-sparsity also leads to inaccurate user features learning in collaborative filtering recommendation algorithms.To solve this problem,this thesis proposes an academic paper recommendation model UP-PMF that collaboratively learn the user preference and user ratings.Meanwhile,a corresponding multitask learning model is also proposed.Firstly,a user preference model based on self-attention mechanism is constructed to obtain user features from the content of the papers read by users.Then,the model is used to constrain the user latent feature vectors in HAN-PMF to obtain more accurate deep feature representation of users.Finally,by adding the tag prediction task,the corresponding multi-task learning model is realized.The results of the comparative experiment show that the model achieves better performance than the current similar recommendation model.