Study And Implementation Of Recommender System Based On Deep Learning

Information overload has become one of the major challenges in the development of Internet applications.In the era of big data,the Recommender System(RS)is the most common solution to alleviate information overload,and has been widely adopted by websites and applications.The deep learning method is the hot spot for the research of recommender system.However,due to the lack of deep understanding of the feature extraction of text and user data in the recommender system,most of the existing text-oriented depth recommendation algorithms do not fully exploit the text and user feature information,and the prediction accuracy is low.Therefore,it is necessary to design a text-oriented deep recommendation algorithm based on the analysis of the characteristics of text data and user data in specific scenarios like websites or applications that use text as additional feedback for recommendation(such as Taobao and today's headlines).Based on the starting point,this paper proposes a recommendation algorithm based on attention mechanism for text data and user data.And then this paper proposes a hybrid prediction layer for mixed data.Finally,this paper designs and implements a newsletter system.The specific work includes:Firstly,this paper proposes a multi-level attention mechanism in the deep recommender system.The mechanism is able to train attention weights at the word level and full text level.It automatically selects words that are closely related to recommendations,filters low-quality reviews and assigns different weights to different reviews,and makes more accurate predictions of user ratings.In addition,the attention mechanism provides a commentary basis for the interpretability of the recommendation results,which can filter out comments and keywords that are of interest to the user,and help the user to get rid of the way to obtain the item information by reading a large number of comments.This has great practical significance for the current recommender system.(corresponding to Chapter 2)Secondly,this paper proposes the hybrid prediction layer which couples factor factorization machine and the multilayer perceptron to model the complex nonlinear relationship between users and items.The factorization machine implements the modeling of low-order interactions;the multi-layer perceptron implements the modeling of nonlinear interactions.Although there are many hybrid models that combine traditional and deep methods,the current mainstream hybrid model is primarily for click-through rate prediction.The hybrid model can realize a mixed prediction structure with weight sharing for scoring prediction,achieve good results,and provide some reference significance for related research.(corresponding to Chapter 3)Thirdly,the paper designs and implements a news recommender system.The system combines two algorithms mentioned above to achieve an end-to-end depth framework.The core function is a news portal that can be personalized and recommended.The system includes front and rear ends to enable personalized newsletter.In addition,the system has functions such as user registration,news grab,text pre-processing,personalized recommendation,newsletter,unsubscribe,etc.,which are closely related to news recommendations.(corresponding to Chapter 4)...