Research On Recommendation Algorithm Based On Heterogeneous Information Network And Deep Learning

With the rapid development of the Internet era,users are overwhelmed by the amount of information on the Internet.In order to alleviate the problem of data overload and help users quickly find products or services that may be of interest,recommendation systems have been proposed and played an important role in various online platforms.Traditional recommendation filtering algorithms usually use the user’s historical interaction records to model user preferences.However,the score matrix is usually sparse,which limits the performance of such algorithms.As the information of users and items in the recommendation system gradually increases,researchers have begun to try to use auxiliary information to alleviate the problem of data sparseness in the recommendation system.Due to the flexibility of heterogeneous information networks in processing multiple types of data,they have been used to model the rich information in recommendation systems in recent years.At the same time,due to the strong nonlinear modeling capabilities of deep learning,they have also been widely used in the field of recommendation systems.Therefore,this thesis attempts to combine heterogeneous information networks and deep learning techniques to improve recommendation performance.The main tasks are as follows:This thesis proposes a neural recommendation model Trust NCF that integrates multiple trust relationships.Considering that users are often affected by trusted users when choosing items,Trust NCF improves recommendation performance by adding trust relationships between users.The user-item interactive information and the trust relationship between users can be jointly constructed as a simple heterogeneous information network.With the spread of social information and interactive information in the network layer by layer,the user’s interest continues in the recursive process changing.In this thesis,a concept of a hierarchical trust relationship is proposed,and a communication influence layer is designed,which includes a trust communication layer and an interactive communication layer.The interaction and trust relationship between nodes are respectively transmitted,and the neighbor relationship is realized through the graph convolutional network.The multi-layer aggregation of domain nodes simulates the propagation process of node influence in the network.We verify the Trust NCF model on the Epinions and Last FM datasets,using NGCF,Precision and Recall as evaluation indicators.Compared with the traditional social-based recommendation and graph-based recommendation,the performance of the Trust NCF model is significantly improved.This thesis proposes a recommendation model MLFNNR that integrates multiple potential features and neural networks.Considering that users and items usually have a large amount of attribute information,MLFNNR uses this type of attribute information to improve recommendation performance.The interaction information and attribute information between users and items can be modeled as a heterogeneous information network containing rich semantics.First of all,this thesis proposes the use of meta-paths to extract the node sequence that contains different aspects of information in the network,and then learns the potential feature representation of the nodes through the heterogeneous skip-gram model.After obtaining various features,combine the attention mechanism to weight and fuse these feature vectors,and then consider that the neural network can replace the inner product form of the traditional matrix factorization,and use a nonlinear function to simulate the interaction between the user and the item to achieve Score prediction.The MLFNNR model was evaluated on the Movielens and Douban movie datasets,using MAE and RMSE as evaluation indicators.Compared with traditional methods and methods based on heterogeneous information networks,the MLFNNR model can significantly improve recommendation performance,and at the same time,we also verified the multi-faceted features and the necessity of using neural structures through experiments.