Research And Application Of Personalized Recommendation Algorithm Based On User Community

With the development of information technology and the popularization of ecommerce,more and more data need to be processed every day.Individuals can't obtain the information they need as efficiently and accurately as before.Therefore,the importance of the recommendation systems which are designed for the massive information filtering is constantly increasing.At the same time,people's requirements for the recommendation systems are getting higher and higher,they hope that the system can filter the information they need out of the data in time.In order to achieve this goal,the researchers introduce the user community concept into the field of recommender systems.Community detection algorithms in which similar user communities are detected and the scope of calculations when recommending is reduced.In order to increase the efficiency and quality of the recommender systems,this thesis mainly studies on the communitybased personalized recommendation systems.Firstly,a personalized recommendation algorithm based on non-overlapping communities of users is proposed in this thesis.Before user communities are detected,user features dimension is reduced by SVD and high-dimensional sparse features of users are mapped to a low-latitude feature space.Then non-overlapping communities are detected by K-Means++ algorithm and items are recommended to the users in the same community.The experimental results show that this new algorithm performs better than the traditional recommendation algorithms.Secondly,another personalized recommendation algorithm based on overlapping communities of users is proposed.The overlap attribution of user communities is measured by Fuzzy C-Means clustering algorithm.The core users of each community are chosen by membership and items are recommended to users in this community by UCF.After all communities are handled,the final recommendation list is provided by the algorithm using membership to hybrid.The result of experiments shows that this new algorithm achieves higher precision and uses less time compared to the traditional recommendation algorithms.Finally,considering user features dynamic updating and new users entering the system,an algorithm based on incremental matrix decomposition and matrix projection is proposed in this thesis.The algorithm makes the recommendation model possible to handle user's feedback and reflect the feedback in the recommendation results in time.The scalability of the recommendation model and the real-time computing capability are improved.The projection operation projects the user features of the new user onto the target user features space and make it possible to provide the new user a better recommendation list than the GRM algorithm.In the end,it is found that this dynamic updating algorithm shows good results on different data sets.