Research On Collaborative Filtering Recommendation Methods Based On Deep Learning And Heterogeneous Information Networks

With the continuous development of mobile Internet and various intelligent terminal devices,the information on the network has exploded.In order to alleviate the problem of information overload,the researchers proposed to build a recommendation system to provide users with personalized information,products or services to meet their preferences.However,The problem of scoring data sparsity limits the recommendation performance of traditional collaborative filtering recommendation algorithms to a certain extent.With more and more information about users and articles,how to make full use of this information to break through the performance bottleneck caused by the sparseness of scoring data has become one of the hot issues in the recommendation system.The existing work pays less attention to how to combine heterogeneous information networks and deep learning to achieve better recommendations.This paper proposes deep collaborative filtering recommendation models based on entity-based and path-based heterogeneous information network information.The main work is as follows:1.This paper proposes the TrustDMF model.TrustDMF allows users to extract loworder features from the level of the item,and the item and the user's trusted user group extract low-order features from the user's level.Then,the high-order features are extracted by MLP.Finally,the cosine similarity of the feature vector of the user and the item is used as the predicted score value,and the cosine similarity of the feature vector of the user and the user's trusted user group is used as the predictive trust degree.for a good consideration of trust information,the loss function increases the social regularization constraint,which increases the squared difference between the prediction trust and the actual trust.2.This paper proposes the PANCF model.PANCF uses LSTM to model the path to extract the feature vector,then uses the max-pooling operation on the feature vector group of the path belonging to the same meta-path to obtain the feature vector of the meta-path,and then fuses the feature vectors of all the meta-paths with the attention mechanism to get the feature vector of the context of the user-item pair.Finally,the final prediction is made by using the fully connected layer after splicing the feature vector of the user and the item and the feature vector of the context information.3.In this paper,the TrustDMF model is validated on the FilmTrust and Epinions datasets.The results show that the TrustDMF model is 1.2% and 1.7% higher than the DMF model on HR and NDCG,respectively.The PANCF model was experimentally verified on the Movielens and LastFM datasets.The results show that the PANCF model has improved by 1.8%,1%,and 1.2% compared with the MCRec model in Precision,Recall,and NDCG.