| With the development of information science and technology,a large amount of information is gathered on the Internet,which also brings about the problem of information overload.How to accurately recommend to users the content they are interested in becomes more and more important.Faced with this problem,the importance of recommender systems is selfevident,especially as users face more and more content and make choices more and more difficult.This Thesis studies matching and ranking stage in recommendation systems,and models users’ long and short-term interests,and mainly completes the following three tasks:(1)Aiming at the modeling of users’ long and short-term interests in the matching stage,a LSLM matching algorithm based on LHUC and AMSoftmax is proposed.Based on the factorization machines(FM)network structure,LSLM allows users to cross each other’s static features,mines and utilizes rich user portrait information,and details the long-term interests of users.For the user’s short-term interest,the Transformer structure is used for modeling,and LHUC is introduced into the neural network to model the user’s personalized bias.At the same time,the gating mechanism is used to dynamically fuse the user’s short-term interest and user’s long-term interest.By leveraging the AM-Softmax loss function,LSLM achieves classification and ranking consistency.LSLM achieves 0.5465 and 0.1468 on HR@50 and MRR on the MovieLens dataset.The experimental results show that the LSTM model has greater user interest coverage than other matching models.(2)A BERTMF model based on BERT and generalized matrix factorization is proposed for modeling users’ long-term and short-term interests in the ranking stage.BERTMF models users’ long-term interests through a generalized matrix factorization model.GMF is based on the global co-occurrence of users and items,which enables the model to well represent users’ long-term preferences.BERTMF uses the BERT model to model the user’s recent behavior sequence,and obtains the user’s current preference expression through the mask mechanism.In order to avoid the divergence and drift of users’ short-term interests,a fusion layer and an attention layer are designed to model the user’s fusion interest preference.BERTMF achieves a high AUC of 0.7714 on the Electronics dataset.The experimental results show that the AUC of BERTMF is higher than other models on multiple public datasets.(3)Based on the Django framework,a user-personalized movie recommendation system is constructed.The system uses LSLM algorithm and BERTMF algorithm as the basic model for personalized recommendation. |
