Font Size: a A A

Research On Hybrid Recommendation Algorithm Based On Convolutional Neural Network

Posted on:2021-01-16Degree:MasterType:Thesis
Country:ChinaCandidate:Y CuiFull Text:PDF
GTID:2428330626965641Subject:Engineering
Abstract/Summary:PDF Full Text Request
With the rapid development of network environment in recent years,network information is spreading to all the aspects we need.The data that people get online is becoming more and more rich,but this situation has led to the rapid growth of data volume.According to the results of statistics,the active Facebook users share about684,000 bits of information per minute on the web,while Twitter users issue more than 100,000 tweets.90% of the world's data emerged in 2010-2012,by 2020,the global information will be 22 times what it was in 2011,and total 35.2 ZB.But many of them are redundant data,which led to the problem of "information overload",then the online world surrounded by information.Jack ma,the ex-CEO of Alibaba group,published his new point of view for the network information technology in a speech.He said humans are marching towards the DT(data technology)era from the quondam IT(information technology)era.The IT era is based on self-control and self-management,while the DT era is based on technologies that serve the public and promote productivity.Therefore,under such a rapid increase in the amount of data,the recommendation system came into being and became a necessary tool to help users obtain effective information.As a filtering technology to solve the information overload,it played an important role.The traditional recommendation algorithm is to map user information and item information into a matrix form,calculate their cosine similarity or Pearson similarity function,find out the items of interest to the user through analysis and make recommendations,but in practical applications,the problem of data sparsity caused by incomplete information in the past,the problem of scalability in the case of increased data,and the problem of cold start without previous user information will affect the accuracy of the recommendation results.In view of a series of problems in the recommendation algorithm,this paper focuses on the study of data sparsity and the situation of insufficient user information,improves the traditional recommendation algorithm,and proposes a hybrid recommendation algorithm.After long-term online reading and detailed analysis of the online book store,it is found that at this stage,the Internet's attention and recommendation accuracy for books needs to be improved,and readers cannot accurately find books with similar interests among many books on the Internet and can't find similar books that have been read in the past,so it takes a lot of time for the readers to find.So,this article applies the improved algorithm to the book data set.Toexplore the huge number and variety of books,the algorithm is applied to the book data set for verification.The main tasks are as follows:1.Through research on the relevant theoretical knowledge,technology,background of the recommendation system,the author deeply understand the far-reaching impact of the recommendation system on the future development of network applications;summing up the shortcomings of traditional recommendation algorithms based on the comparative research on traditional algorithms,the author decided to put solving Data sparse problems and the analysis of recommendation algorithms integrated with sentiment polarity as the focus of this paper.2.Aiming at the problem of data sparseness,the data preprocessing method is used to remove users who have not rated any items and items that have not been rated by users.Then,according to the past behavior of users,the relationship between users and the relationship between users and items are mined to pass the rated items.This paper focuses on building user pairs and item pairs from the scored items,calculating the similarity,constructing a similarity matrix,then form the nearest neighbor set,and predict the score through collaborative filtering;integrate the XGBoost algorithm to classify users and items,calculate the classification error rate,and update the weight and learning rate by training the algorithm,find the wrong sample,and reset the weight to achieve accurate classification.3.The research found that the deep-level features of the user's emotional information have a greater impact on the accuracy of the recommendation algorithm,so this article models the user's evaluation information to find that it contains emotional information,judges the emotional polarity,and uses the standard library to roughly conduct user interests classification,and then train the convolutional neural network according to the user,project information,user prediction score for the project,and adjust the learning rate,the number of convolution kernels and other related parameters through back propagation.To train the final score,then rank the scores,and recommend the item to the user according to them.4.In order to prove the effectiveness of the algorithm proposed in this paper,the author conducted an experiment on the public data set of Douban reading top250 book information and popular reviews collated by Tsinghua University.The data set was first cleaned,and the data that did not make any comments on the project were carried out.Through clearing,not only reduces information overload,but also reduces the amount of data,which is beneficial to the later training process.Then,theXGBoost algorithm was used to classify,further solving the "data overload" situation.In the later scoring prediction,the accuracy is improved,and the training time is significantly reduced.Experimental results show that the algorithm proposed in this paper has a certain improvement in recommendation accuracy compared with the comparison algorithm,and has certain application value in the field of book recommendation.
Keywords/Search Tags:Collaborative filtering, XGBoost, Affective polar, Convolutional neural network, Accuracy
PDF Full Text Request
Related items