Recommendation Technology And Application Based On Tag Transfer Learning

With the rapid development of information technology and the gradual realization of economic global integration,the Internet has entered an era of rapid development.The Internet Trends report released at the Code conference in May 2018 shows that Internet users have experienced explosive growth in recent years.In 2017,the global Internet penetration rate reached 49%,and is expected to reach 50% in 2018.According to statistics,after the global big data enters an accelerated development period,the total amount of global data will increase with a rate of 50% per year.Meanwhile,the global data volume will increase to 176 ZB in 2025.Faced with such massive data information,it is one of the most important challenges to alleviate the problem of information overload,in order to help users to obtain "useful" information quickly and efficiently.Recommender Systems(RS)is one of the effective ways to solve the problem caused by information overload.By mining the binary relationship between users and projects,RS can help users discover the items they may be interested in from a large amount of data and generate personalized recommendations to meet individual needs.Tag-based Recommendation Technology provides users with personalized recommendation by utilizing user's tag information on the project.However,the existing Tag-based Recommendation methods are still subject to the following challenges and limitations related to data characteristics: 1)when personalized recommendation is made for users,it tends to give more attention to popular tags,which leads to weight deviation and reduces the novelty and accuracy of recommendation results;2)training samples and new test samples for learning need to satisfy the independent and identically distributed strips.3)The labeled sample data sets are usually very sparse and difficult to obtain,and sufficient training samples must be available to learn a good recommendation model.In view of the above problems,the main contributions of this paper are as follows:1.To solve the problem of weight deviation,a collaborative filtering personalized recommendation algorithm based on label entropy feature representation is proposed.Firstly,the label information entropy is used to measure the uncertainty of the label and “punish” the hot label.Secondly,the user-labelitem relationship is described by three-part graph.The user-item feature representation based on label entropy is constructed,and the project similarity is calculated by feature similarity measurement method.Finally,the user's preference value is predicted by using the linear combination of user label behavior and item similarity,and the final recommendation list can be generated by ranking the predicted preference values.The experimental results on Last.fm dataset show that this method can improve the accuracy and novelty of recommendation and meet the personalized needs of users.2.Aiming at the problem that training samples and new test samples for learning can not satisfy the assumption of independent,identical distribution and data sparsity,a recommendation algorithm based on label feature information transfer learning is proposed,which transfers user and item feature factors learned from source domain data to target domain.Firstly,the label is used to construct the feature representation of the project in the auxiliary data set,and the user's feature representation is calculated according to the behavior data of the user's selection of the project set.Then,the characteristics of the project are measured according to the user group of the selected project in the target data set,and then the features of the user and the project are smoothed in the target data set to eliminate the influence of different scoring scales of the data set.Finally,user and project features are applied to collaborative filtering personalized recommendation algorithm based on label entropy feature representation,and a recommendation method based on label feature information transfer learning is designed and constructed to complete target project recommendation.The experimental results on Movie Lens dataset show that this method can effectively alleviate the problem of data sparsity and improve the accuracy of recommendation.3.Based on the proposed recommendation algorithm,this paper designs and implements a music recommendation application system.Based on Spring,My Batis and Spring MVC frameworks,Java,Python,HTML and JQuery are used as development languages.User data is computed using Python encapsulated interface.Tags are vectorized.Feature representation and project similarity are constructed according to the historical records of users and projects using tags,in order to recommend music they may like for the target users.By integrating Python in Java and calling Python script files,the recommended results are dynamically returned to users.Users of the system only need to provide the user ID of the target user.The backstage of the system will generate Top-5 music recommendation for users according to their historical behavioral data.