Personalized Recommendation Based On Background Information And Autoencoder

The proliferation of Internet data brings about the problem of "information overload",and personalized recommendation is an effective means to alleviate this problem.Personalized recommendation aims to recommend products or services that may be of interest to users based on their past behavior.Collaborative filtering has become a promising method due to its effectiveness and robustness.Traditional collaborative filtering uses matrix factorization methods to learn the hidden feature representations of users and/or items,but matrix factorization methods often only learn linear features and perform poorly when the data is sparse.In recent years,deep learning has achieved good performance in representation learning.As an excellent deep learning model,the autoencoder model is widely applied in recommendation systems for the advantages of fast convergence and no label requirement.In addition,recommendation methods based on autoencoders often use additional attributes to solve the problem of data sparsity.However,traditional recommendation based on autoencoders using additional attributes to alleviate data sparsity has two main disadvantages.The first is the high computational cost of most existing recommendation models when using additional information from users and/or items to expand the feature space.The second disadvantage is that it is difficult to obtain additional user information due to the high cost of acquiring tag knowledge and the increase in user privacy awareness.This dissertation focuses on the research of personalized recommendation algorithm based on background information and autoencoders for these two problems.The main research contributions are as follows.(1)A HCRDa model is proposed in this thesis.This model uses two semi-autoencoders to co-embed the rating vectors and attribute features of users and items to learn the hidden features of users and items at the same time,and combines the matrix factorization to achieve personalized recommendations.Firstly,two novel semi-autoencoders are utilized to simultaneously learn the feature representations of users and items in our HCRDa.Secondly,embedding the matrix factorization technique into the training process of the autoencoder further improves the quality of hidden features for users and items.Finally,the additional attributes of users and items are utilized to alleviate the cold start problem and to make hybrid recommendations.(2)This thesis proposes a Co-Agpre method,which uses a semi-autoencoder to coembed the rating vectors,attribute features and graph features of items for rating prediction.More specifically,a semi-autoencoder is introduced to learn the hidden nonlinear features of items for achieving a low computational cost,and thus the proposed Co-Agpre model can flexibly use side information from different sources.Meanwhile,in the case that it is not easy to obtain the user's additional attributes,we take the item's graph features and attributes into consideration for improving the accuracy of recommendation.Experiments on real-world datasets demonstrate the effectiveness of the proposed Co-Agpre compared with state-of-the-art attribute-aware and content-aware methods.