Research On Personalized Recommendation Algorithm Based On Users’ Implicit Information Mining

With the rapid development of the Internet,massive online information has brought great convenience to people’s daily life.However,too much available online information also makes it difficult for people to find the knowledge and information they are interested in.Towards this problem,personalized recommendation algorithms came into being.These algorithms are developed to provide personal recommendations for each user to feed their interests as much as possible by modeling users’ or items’ preferences.At present,the recommendation algorithm has been successfully applied to a large number of online applications,such as Taobao,Weibo,Tik Tok and so on.Their success makes it become a hot research topic to develop recommendation algorithms for both commercial and research value.Although researchers have put forward many effective recommendation algorithms,the data sparsity problem is still one of the key problems that restrict the improvement of recommendation algorithms.For that users are usually only interested in a small number of items and leave few feedback information on items,there are few the feedback data that can be used to directly estimate each user’s preferences.This problem challenges the accuracy of recommendation algorithms.Therefore,in order to mine more users’ implicit information from limited data to solve the problem of data sparsity,this thesis studies the relationships between each pair of users,between users’ multiple roles,and between each pair of user and item in many aspects,and then achieves the corresponding research results.This work provides solutions to solve the problem of how to better mine more information from existing data.So that this thesis has certain practical and scientific significance.The main contributions of this thesis are listed as follows:(1)We study the problem of the influence of multi-role information on trust strength.For that each user trust others with different degrees,so it is necessary to study the problem of how to estimate the trust strength between users.However,the existing methods do not consider the influence of users’ multiple roles on trust strength calculation.To solve this problem,this thesis proposes a trust strength calculation method by considering users’ multiple roles for recommendation.This method is developed to estimate and calculate the trust strength between users from multiple perspectives,so as to measure the trust influence between users more accurately.(2)We study the problem of the information of direction problem in trust propagation.The trust relationships between users are directional,and the existing related works have the following shortcomings.First,they ignore direction information when modeling the influence of trust propagation between each pair of truster and trustee.Second,the values of trust strength are not constrained to be positive number in their objective function,which may lead to unreasonable results.In order to solve these problems,we conduct in-depth exploration from the perspective of directed trust propagation,and then propose a personalized recommendation algorithm.(3)We study the problem of the relevance among users’ multiple roles.The existing work focuses on how to model the multiple roles for users,but ignores the correlation among the multiple roles of users.In order to mine more hidden information of users,this thesis proposes a correlative denoising autoencoder algorithm based on deep neural networks to learn the correlations between users’ multiple roles for recommendation.(4)We study the problem of information completion of users’ data.It is helpful to improve the accuracy of the recommendation algorithm by completing the missing data of users through the pretraining model.However,the existing related work still has the following problems.First,they adopt the phased training method which is not conducive to transfer the information of the back-end module to the front-end module for accurate improvement.Second,they directly use the generated data for retraining,while ignore the noise caused by imprect pretraining model.In order to solve these problems,we propose an enhanced collaborative autoencoder algorithm,which integrates complementary information for recommendation based on deep neural networks.