A Research Of Cross-domain Recommendation Algorithm Based On Tag Mapping And Transfer Learning

With the rapid development of computer hardware and software technology,the Internet has entered the era of Web 2.0,and Internet applications have been born in large quantities.At the same time,a huge amount of unstructured data has been generated.How to mine the potential value of these data has become the research content of more and more scholars.In some data-driven applications,the problem of ”information overload” often occurs.For example,in the field of e-commerce,users face a huge list of goods,how to make choices,how to obtain the most effective information,and what information enterprises should show to users,and whether the information presented is of interest to users.These problems often bring troubles to both sides.In order to solve these challenges,recommendation system came into being.The basic core of the recommendation system is to mine the internal association and potential features from the user information,the item information and the interaction information between the user and the item,so as to contact the user and the item,so as to meet the specific industrial needs such as recommending the item for the user.In the traditional single domain recommendation algorithm,the neighborhood based algorithm is first proposed,but it only considers the interaction information between users and objects,which is insufficient in many scenarios,and it is difficult to deal with the data sparsity caused by the growing data.Therefore,it is necessary to mine more relevant additional information and build a more elegant model to improve the recommendation effect.The recommendation method combining the attributes of goods and users has been proved to be effective.This thesis will study how to improve the recommendation effect from the dimension of social tags.In the cross domain recommendation system,another challenge of the recommendation system is the cold start problem.How to capture the user's interest quickly is a difficult problem to be solved.This thesis will explore the user's cold start problem in the new domain from the perspective of user's cold start and migration learning.On the basis of reading the related literature of cross domain recommendation,this thesis summarizes two common problems in cross domain recommendation: how to migrate the tags in the source domain and the target domain;how to predict the users' ratings of the items in the target domain,and makes the question definition and question analysis with mathematical symbols.To solve the problem of tag data migration,this thesis uses tag data as a bridge between source domain and target domain,uses improved tag cooccurrence technology to get the vectorization representation of tags,and then proposes tag embedding cluster(TEC)and tag transfer(TT),generate the topic model,then migrate the tags of the target domain to their respective topics,and finally get the tag topic probability matrix(BCP)for scoring calculation.To solve the problem of cross domain user rating,this thesis proposes a topic factorization machine(TFM)model based on multilayer perceptron,which takes the combined characteristics of users and tags as input,and then users get the user topic preference matrix(UCP),and finally predicts the user's rating of objects in the target domain according to BCP and UCP.In this thesis,we use the movielens dataset and amazon dataset with label data and scoring data.First,we clean,construct and process the data set according to the experimental requirements.Then we design experiments to study the effectiveness of label migration clustering algorithm.Through the correlation dimension reduction method,we carry out visual analysis on the experimental data,and the experimental results are in line with expectations.For the subject factor decomposition machine model,we first Observe the experimental results under different parameters,analyze the super parameters which have great influence on the results,then design the parameter sensitivity experiment,analyze and select the parameter value to obtain the optimal solution.Finally,this thesis selects the basic and the latest correlation comparison method,designs the comparison experiment under different label overlap,and the experiment proves that the model proposed in this thesis is under the specific parameters It is better than the comparison method and improves the recommendation effect to a certain extent.