Personalized Recommendation Based On Transfer Learning

With the development of computer technology and the rapid expansion of the scale of the Internet,people can easily access to more and more network information.However,it is difficult for people to quickly find the information they need when they face the massive information.The increase of information leads to the decrease of people's efficiency of obtaining information they need,which is the problem of information overload.Therefore,how to help users effectively,quickly and accurately obtain the information or knowledge they need in such a huge and complex information,that is,how to alleviate the problem of information overload,has received more and more attention and research.At present,personalized recommendation algorithm is an important method to alleviate the problem of information overload.It analyzes the various attribute information of users and items and the various interactive data between users and commodities,mining the potential preference information of users to accurately predict the user's preferences,and provides per-sonalized recommendation services for users,so that users can quickly and effectively find the goods or services they need in a large number of goods.Personalized recommendation algorithm has become an independent research direction,and it has experienced continuous development,new recommendation methods are emerging in endlessly,and its accuracy is constantly improved.However,although personalized recommendation algorithm is be-coming more and more mature,there are still some problems to be solved,the most serious problem is data sparsity.The lack of data has a great influence on the accuracy of personal-ized recommendation algorithm.In order to solve the problem of data sparsity in personalized recommendation algorithm,some scholars put forward the idea of transfer learning applied to personalized recommen-dation algorithm.The main idea of transfer learning is as follows:firstly,find the target domain related and data dense domain as the auxiliary domain,and the data in the auxiliary domain as the auxiliary data,then extract the shared key knowledge and information from the auxiliary data with appropriate methods,then transfer the knowledge and information to the target domain in an appropriate way.The knowledge transferred from auxiliary domain could help train the data in the target field.The idea of transfer learning applied in person-alized recommendation algorithm is to take the score data as the target data,take the data which is closely related to it and may be helpful for its prediction(such as the score data in other fields)as the auxiliary data,and then use appropriate methods to select the auxiliary data from the auxiliary data.The key knowledge and information shared by the two data sources are extracted and transferred to the target domain to help the recommendation algo-rithm do better prediction and improve the accuracy of recommendation.A large number of practices have proved that transfer learning is an effective method to alleviate the problem of data sparsity in the process of personalized recommendation.However,there are still some problems and deficiencies in the existing personalized recom-mendation algorithms combined with the idea of transfer learning,which need to be further improved.At present,personalized recommendation algorithm based on transfer learning mainly faces two problems:first,when using transfer learning for cross domain recommen-dation,current models and algorithms seldom consider the differences between the two do-mains,which may lead to negative data migration and reduce the recommendation accuracy.Secondly,the process of current models and algorithms in practical application is generally to find the data intensive auxiliary domain first,and then do knowledge transfer and cross domain recommendation.However,in reality,the data intensive auxiliary domain similar to the target domain may not be found,and many personalized algorithms ignore the im-portance of adjacent users in recommendation.Therefore,how to use the idea of transfer learning to make personalized recommendation in a single domain and consider the influ-ence of neighboring users is a challenge we are facing.In this paper,the above two problems are studied.Specifically,the main work of this paper is as follows:(a)This paper proposes a cross domain recommendation model called Knowledge Sepa-rated Codebook Transfer(KSCBT)based on latent feature clustering.KSCBT model is an improvement of Codebook Transfer(CBT)model.CBT model first compresses the ratings in the auxiliary rating matrix into a cluster level rating pattern representation with rich and compact information,which is called codebook.Then,CBT proposes an effective algorithm to reconstruct the target evaluation matrix by extending codebook,which transfers the scor-ing mode of auxiliary domain to the target domain.However,CBT model does not fully consider the differences between different fields.Based on CBT,KSCBT model proposed in this paper not only learns"shared knowledge" between different fields,but also"special knowledge" between different fields."Special knowledge" represents the part of"knowl-edge" which only belongs to various fields.This part of"special knowledge" can alleviate the over fitting problem in transfer learning to a certain extent.The specific process of KSCBT is as follows:firstly,an Adaptive Codebook Transfer Learn-ing(ACTL)model is used to obtain a codebook of k×l size suitable for the source domain.Then,CBT is used to construct a codebook with the same size of k x I in the target do-main.Then,Dynamic Programming(DP)distance matching algorithm is used to calculate the similarity between the dimensions of the two codebooks,and the similarity is used to de-termine which dimensions are public knowledge that can be transfered and which are private knowledge that can be retained.The experimental dataset of this model adopts three open datasets used in recommendation systems:MovieLens dataset,EachMovie dataset and Book-Crossing dataset.Movielens dataset is used as auxiliary domain,EachMovie dataset and Book Crossing dataset are used as target domain respectively.During the experiment,we use KSCBT model to transfer the public knowledge learned from MovieLens dataset to EachMovie and Book-Crossing,while retaining the private knowledge of the target domain.The experimental results show that the MAE value of KSCBT model is smaller than MAE of other baseline models including CBT,which indicates that KSCBT model has higher recommended accuracy.(b)This paper proposes a Single-domain Recommendation based on Auxiliary-domain Con-structed and KSCBT(SRACK).In this model,we propose the auxiliary data construction algorithm called Binary-preference Auxiliary Data Construction(B-ADC),Similarity Cal-culation Based on Sparse Data(SCBSD),Factorization Transfer based on Neighbors(FTN),and combine them with KSCBT.The process of SRACK as follows:We construct the binary preference auxiliary data from the original scoring data to create the auxiliary data,which avoids the problem that the low correlation between the auxiliary data and the target scoring data leads to the decline of recommendation performance.Then we propose a similarity calculation algorithm called Similarity Calculation Based on Sparse Data(SCBSD).We use SCBSD to calculate the sim-ilarity of users.On the basis,we propose a personalized recommendation algorithm called Factorization Transfer based on Neighbors(FTN)based on the idea of matrix factorization and transfer learning.FTN algorithm not only migrates users' binary rating preference fea-tures to rating preference features,but also considers the influence of neighbor users on users' scoring.Specifically,FTN migrates the binary rating preference features of the near-est neighbor users to the rating preference features of the nearest neighbor users,then use the new rating preference feature and new new rating preference feature of neighbors to do rating prediction.When get the adjacent neighbor users,FTN uses SCBSD.SRACK uses B-ADC and FTN to predict the vacancy value in the target domain,but this paper does not directly take the prediction result as the final result,but takes the prediction result as the auxiliary domain,and uses KSCBT model to process the final result.This is equivalent to that the target domain uses its own data,constructs an auxiliary domain with B-ADC and FTN,and then uses the auxiliary domain for"cross domain recommendation".The experimental datase of this model adopts the open data sets used in two recommen-dation systems:MovieLens dataset and EachMovie dataset.SRACK model is used to train and predict the two target domains.The experimental results show that the MAE value of SRACK model is smaller than that of other single domain personalized recommendation baseline models,which indicates that SRACK model has better recommendation perfor-mance.The work of this paper is summarized as follows:KSCBT and SRACK are two new models proposed in this paper.KSCBT aims at solving the problem of negative transfer by using the method of distinguishing private knowledge and public knowledge when using the idea of transfer learning for cross domain recommendation;SRACK model aims at solving the problem of negative transfer when there is no way to find suitable auxiliary domain to use the idea of transfer learning for cross domain recommendation The experimental results also show that SRACK is superior to some single domain personalized recommendation algorithms and some cross domain recommendation algorithms.In the future work,for KSCBT model,when obtaining codebook,ACTL algorithm can only obtain the codebook size most suitable for the source domain,which is not necessarily the most suitable for the target domain,so how to balance the codebook size is a challenge in the future;for SRACK,its complexity is high,and whether it has excellent performance in the context of big data needs to be tested.