Research On Personalized Recommendation Based On Double Attentional Deformable Convolutional Network

The rapid development of information technology not only brings great convenience to people,but also brings explosive exponential growth of network information.In order to alleviate the problem of information overload,recommender systems came into being,developed rapidly and have been successfully applied in many fields.However,due to the sparsity of rating data and the limitations of traditional recommendation algorithms,it is difficult to further improve the recommendation performance.In recent years,review text with rich semantic information and deep learning technologies with powerful data processing ability have brought new opportunities to alleviate the problem of data sparsity.In order to improve the performance of recommendation,we use the review information,the deformable convolutional network and the attention mechanism to improve the recommendation method.This paper mainly includes the following three parts:1.Aiming at the inherent geometric limitation of traditional convolutional neural network and the limitation of factorization machine in feature interaction,a Deformable Convolution and Hybrid Prediction based joint deep Network(DCHPN)recommendation model is proposed.The model is composed of two parallel deformable convolutional networks,which can jointly learn the deep feature representations of users and items freely and flexibly from reviews reflecting user preferences and item attributes.In the hybrid prediction module,the deep neural network and factorization machine are integrated to realize the interaction of low-order features and high-order features.By comparing the experimental results on three different real-world datasets with the performance of baselines,the effectiveness of the model DCHPN is verified.2.Aiming at the difference of information quality of different reviews and the important influence of review usefulness on feature extraction,a Double Attention mechanism based Deformable Convolutional Network(DADCN)recommendation model is proposed on the basis of DCHPN.The model adds the word-level attention mechanism and the review-level attention mechanism to parallel deformable convolutional networks,which can assign relatively high attention weights to critical words and informative reviews.The extended latent factor model integrates the representations of users and items learned from the review text with the features of users and items obtained from the rating data to complete the rating prediction.Experimental results on four different real-world datasets show that the performance of the model DADCN is better than that of the baselines and the model DCHPN.The comparisons between variants and the interpretability analysis of the model fully demonstrate the role of each part of DADCN in performance improvement.3.Aiming at the problem of user’s personalization and the sparsity of user’s reviews,an Auxiliary review based Personalized Attentional Deformable Convolutional Network(APADCN)recommendation model is proposed on the basis of DADCN.The model uses ID information to improve the double attention mechanism in DADCN into personalized wordlevel attention mechanism and personalized review-level attention mechanism.By searching for the reviews of like-minded users who share similar interests with the target user,the user auxiliary network extracts the features of user auxiliary reviews,which are combined with the features of users and items learned from parallel networks in the extended latent factor model to complete the rating prediction.The experimental results on four different real-world datasets show that the performance of the model APADCN is better than that of the baselines and the model DADCN,and the effective improvement of recommendation performance is achieved.