Research On Text Content Recommendation Based On Attention Neural Network

In the current big data environment,as the scale of text applications increasing,the challenge of information overload is getting more and more obvious,which has become a problem that urgently need to be solved in the field of text applications.The recommendation system,as a special means to provide decision support to users based on their information needs and interest preferences,has significant significance for solving information overload.For recommendation methods,good feature extraction capabilities can do more with less.Therefore,The recommendation method based on deep learning with the deep learning's natural advantages in feature extraction capabilities has become an important research hotspot to solve the information overload problem.In the field of text applications,the problem of data sparsity caused by increasing data in the text applications remains serious.And in the past text recommendation tasks,users' long-term interests are often modeled,but through the analysis of the user's psychological regulation,the user's interests will shift with the influence of time,environment,and current state.Therefore,this article starts from the issue of alleviating the data sparseness of the text application platform,and proposes different recommendation models for users' long-term interests and short-term interests.The main research contents are as follows:(1)Aiming at the long-term interests of users,a recommendation model based on multi-channel attention CNN is proposed.The model not only uses explicit data(user information,text that users like),but also uses implicit data(text that users don't like,the author information of text),which enriches the user portrait representation;The sentence vector(PV-DM)method is used to embed the user and text information into low-dimensional dense matrices respectively,which overcomes the problems of sparse coding and high dimensions of traditional text representation methods;Considering that the text content is long,in order to effectively capture the dependency relationship between words and words,words and sentences,attention CNN is used in the text information channel.Finally,the output matrix of each channel is connected end-to-end to obtain the user's final representation,and the Sigmoid function is used to predict the probability of each target text to generate a recommendation.The test results show that the model proposed in this paper is certain effective.(2)Aiming at short-term interests of users,a recommendation model based on attention RNN is proposed.In order to alleviate data sparseness and accurately model user preferences,not only the text content information is embedded in the model,also the user's relationship network is embedded in the model.Use the sentence vector and TransR methods to embed the above two data into two different matrices.Use the connection of these two matrices as an abstract representation of user interests.In the short-term model construction,in order to fully represent the dynamics of the user's short-term potential preference characteristics,Encoder-Decoder is used as the basic framework of the model.The user's short-term preference behavior is used as the input of the encoder.In order to model the dynamics of user preference changes,the Attention mechanism between the encoder and decoder is used.Finally,the recommendation list is obtained through the Bi-GRU layer,the feedforward layer,and the softmax layer.The experimental results show that the short-term recommendation model proposed in this paper is superior in accuracy and can also recommend new text items to users.