Study On Fusing Attention Mechanism And Autoencoder For Recommender System

Recommender systems is an important tool to solve information overloading problem,which is one of the fatal research direction in the field of data mining and machine learning.At present,personalized recommendation algorithms have extremely important applications in various daily fields such as work,study,entertainment,and business.Traditional methods based on collaborative filtering,such as matrix factorization,mine users' preferences by analyzing users' historical interaction records.In recent years,deep learning has been widely leveraged in many research fields,own to stellar performance on learning feature representation.As an excellent deep technique for feature extraction and data dimensionality reduction,autoencoder has attracted the attention of researchers,and many recommendation algorithm models based on autoencoder have been proposed.With the ever-growing volume of online information,the accompanying problem of data sparse problem and popularity bias problem have become a new challenge for recommender systems.Therefore,in this paper,three models are proposed to solve above problems and challenges based on the existing research on matrix factorization and autoencoder.The main contribution of our paper is as follows:(1)Aiming to the performance of matrix factorization is easily influenced by data sparsity,this paper proposes a novel model which integrates stacked sparse autoencoder into matrix factorization(SSAERec).In SSAERec,the feature extraction module is leveraged to learn user latent feature representation and item latent feature representation from the useritem rating matrix respectively,then the extracted representation is integrated into the singular value decomposition model with bias term to predict user ‘s rating towards to item.Comparing SSAERec structure with traditional recommendation methods and deep learning-based methods demonstrates the effectiveness of proposed model.(2)To enhance the poor accuracy of recommender systems caused by partial autoencoder based models that ignore the difference among features,a personalized model A-SAERec is proposed.Meanwhile,in A-SAERec,the neural collaborative filtering structure is utilized to replace traditional matrix factorization to predict the final ratings.The comparative experiments are conducted on four datasets and results demonstrate the effectiveness of proposed A-SAERec on improving the accuracy of recommender systems.(3)In order to boost the low accuracy caused by the recommendation algorithms based on autoencoder ignore the heterogeneity between user reviews and item reviews when employing text information to perform recommendation,as well as to reduce the popularity bias in model.An asymmetrical hierarchical network fused autoencoder is proposed(ANAE).In ANAE,an asymmetrical attentive module is utilized to learning the user review embedding and item review embedding,respectively.Noise contrastive estimation objective is integrated into ANAE for mitigating the popularity bias.The experiments were carried out on four datasets with different sparsity.The experiment results demonstrate the ANAE can effectively reduce popularity bias and boost the accuracy of recommender systems.