Research On Recommendation Algorithm Based On Hierarchical Attention Mechanism And Transfer Learning

Recommendation system is a specific type of intelligent system,which exploits user's historical behaviors and user's information to make recommendations on items to the users.It saves a lot of time cost for users.It plays an important role in e-commerce,social networking,news recommendations and other applications.In recent years,with the successful application of deep learning in some fields,many experts and scholars introduce deep learning into the recommendation system,which brings new opportunities to the recommendation system.However,when people enjoy the convenience brought by the recommendation system and the recommendation system has great significance to real life,there are still problems in the recommendation system based on deep learning,such as insufficient feature extraction,lack of explanation of the recommendation results,sparse data,etc.,which make the recommendation accuracy not high enough.Therefore,in view of the shortcomings in the existing recommendation algorithms,this paper conducts in-depth research in the aspects of feature extraction,improving the interpretation of recommendation results,solving data sparsity,etc.The main content includes:1)A novel recommender with hierarchical attention-based multi-layer perceptrons and factorization machines.In order to alleviate the problem of insufficient feature extraction and lack of explanatory of recommendation results,this paper first proposes a hierarchical attention model to acquire users' interest adaptively,and further analyzes the different contributions of high-order feature interactions and low-order feature interactions to click-through rate prediction.This enhances the explanatory of recommendation results.Then,considering that high-order feature interactions and low-order feature interactions can both play contributions in CTR prediction,we propose an interaction extractor layer to combine Factorization Machine(FM)and Multi-layer Perceptrons(MLP),in which FM extracts low-order feature interactions and MLP extracts high-order feature interactions.Finally,the linear-based global attention mechanism is used to make the high-order feature interaction play different roles from the low-order feature interaction.Experimental results demonstrate Recommendation algorithm of Hierarchical Attention-based Multi-layer Perceptrons and Factorization Machines improves the accuracy of recommendation and enhance the explainability of recommendation.2)Recommendation algorithm based on transfer learning.In order to alleviate data sparsity and improve recommendation accuracy,a recommendation algorithm based on transfer learning is proposed in this thesis.In this thesis,we first look for a source domain similar to the target domain,and then use convolutional neural network to extract the hidden features of comments in the source domain and the target domain.In addition,In order to calculate the domain loss(feature difference)between the source domain and the target domain,a domain adaptation layer is added to the convolutional neural network.Finally,in the training process,domain loss and classifier loss are minimized to complete the depth feature migration and then the recommendation is completed.The recommendation algorithm based on transfer learning effectively alleviates the problem of data sparsity and the recommendation effect is improved.