Research On Some Key Technologies Of Recommender System

With the rapid development of information technology,Internet has developed from a one-way information provider centered by content,into human-centered,interactive network platform stage,and now it is entering the era of the intelligent and individualized Internet of Things(IoT)represented by mobile Internet.On one hand,this network brings convenience to people,but on the other hand,because of the exponential growth of information in the network,what people need and are interested in is flooded in this ocean of information and people face serious information overload.In order to solve this problem,the related technical tools have also experienced three stages of development: classified catalog,search engine,and current recommender system.Recommendation system has become the most important information-filtering tool after search engine.Its goal is to provide users with personalized products and services or to help users make valuable decisions based on this information.The related recommendation technology has been widely studied in the field of information retrieval,machine learning,and data mining.The recommender system has been applied in many areas,such as micro-blogging,news,books,movies,video,music,restaurants and other products and the online recommender systems of Amazon,Netflix,Youtube etc.have been successfully developed and deployed.These systems does not only have brought a good consumer experience to users but also brought great commercial value for the enterprises.In recent years,with the development of mobile Internet,Internet of things,cloud computing,and artificial intelligence technology,the recommendation system has gradually developed into a widely connected,human centered personalized service recommendation model.In this process,the recommendation system faces many opportunities and challenges,mainly in three aspects: first,how to use a large number of side information about users and items on the Internet to solve the sparsity and cold-start problems of the widely used collaborative filtering algorithms.Second,As a result of the development of pervasive computing,context information becomes another important side information source for improving the experience of the recommendation system.Therefore,how to use this new side information to design a context-aware service recommendation system and realize a truly human centered service recommendation has become an important research area.Third,A large number of users,items and context related side information appear,on the one hand,it can be used to optimize recommendation algorithm and improve the system service;on the other hand,the privacy protection of the information is becoming increasingly urgent.Based on these opportunities and challenges,the main research works and contributions of this dissertation are as follows:(1)Based on the matrix factorization collaborative filtering,a model framework for coupling side information is proposed,and a collaborative filtering algorithm for coupling the property information of items is proposed based on this framework.In order to solve the problems of cold-start and low recommendation accuracy in traditional matrix factorization collaborative filtering algorithm,In this dissertation,we modify the user latent factor and the item latent factor of matrix factorization model from the angle of the latent semantic analysis,and forms a model framework.Then,based on this framework,for solving the sparsity and cold-start problem of items,a matrix factorization collaborative filtering algorithm is proposed by innovatively coupling the COS(Coupled Object similarity)similarity between the property of two items.The algorithm uses item similarity to constrain the latent factor factorization of items,so that the more similar items are,the more similar the latent factors are.The experiments show that the algorithm can significantly improve the accuracy of recommendation system and largely alleviate the cold-start problem of item side.(2)Based on the model framework,a matrix factorization collaborative filtering algorithm for coupling social trust relationship is proposed from the user side.Because people tend to seek advice from friends in social networks before buying products or consumer services;based on this hypothesis,this dissertation uses social trust information to calculate the trust value between friends,and integrates the trust values into the proposed model framework to solve the cold-start of the user side and improve the recommendation performance.In order to achieve the model,In this dissertation,we also proposes the social trust measurement,clique detection algorithm,initialization method of user and item latent factor based on contractive auto-encoder.The experiments show that the proposed algorithm can significantly improve the recommendation performance and alleviate the cold-start of the user side.(3)Based on context related side information,a personalized service recommender system framework and an identity recognition algorithm for smart home are proposed.Context is not only the source of user preferences in this system,but also the basis for implementing personalized recommendation services based on AmI(Ambient Intelligence).Therefore,based on the context of smart home and context computing,we proposes a conceptual framework for smart home personalized service recommender system.Then a demo scenario is designed based on this framework.In order to achieve this demo,a identification algorithm based on gait recognition is proposed,and the training and evaluation of the two voice recognition algorithms(GMM-UBM and I-Vector+PLDA)is carried out.The former uses BP neural network to realize identity matching.In the latter,innovative training methods using semi-correlation text are proposed.(4)Aiming at the privacy protection problem of friend recommender system,a privacy protection solution based on hybrid architecture and commutative encryption function is proposed.This solution first uses the KNN classification algorithm for coarse grained friend screening,and then uses the attribute matching protocol to implement the fine-grained recommendation among friends sharing the same attributes.The paper proves theoretically that the protocol can resist active and passive attacks.In addition,we also use anonymous data for the actual deployment and the experimental results show that our proposed model can,not only realize the fine-grained friend recommendation but also to protect the privacy of users.