Research On Expert Users In Personalized Recommendation

With the development of the Internet, the information explosion era has arrived. People often suffer from a large amount of useless information. The importance of personalized recommendation system is coming up gradually, as it not only can help people filter out mass of useless information, but also can increase the income of businesses through the virtual marketing. Therefore, it has aroused great interest of scholars from industry and academia to study personalized recommendation algorithms. In recent years, with the rapid development of social network, the personalized recommendation algorithm based on social network has become the main research direction in the field of recommendation.Cold-start and data sparsity are the inevitable challenges in the research of personalized recommendation algorithm, which can seriously affect the overall performance of the recommendation algorithm. In the research of personalized recommendation based on the social network, researchers mitigate the adverse effects of cold-start users by adding multiple social connections into the algorithm. And many traditional methods use the explicit relationship between users, such as friends. However, it is difficult for cold-start users to have enough explicit relationship for the algorithm to learn. So, we propose exploiting the implicit relationship (expert) between users for personalized recommendation, in other words using experts who are in the field of interest for a user to learn his preference. The main work is as follows:1)When expert users are labeled in the dataset, we use friends and experts in circle to optimize personalized recommendation algorithm. The circle is made up by users who visit a particular category of items and their social connections. Because when user access the category of items that means he is interested in this category. The algorithm designs the regularization term of friends and experts. Then using the regularization terms constrain the objective function of matrix factorization to improve the accuracy of the recommendation. And an extensive experiment is performed on the Yelp contest real dataset to verify the effectiveness of the algorithm.2)When expert users are not labeled in the dataset, an unsupervised expert discovery method is firstly proposed, which is based on the information dissemination theory and the user’s rating record. Then we will combine the impact of experts on users with the objective function to optimize the personalized recommendation algorithm. Finally, plenty of experiments are performed on Yelp and Epinions datasets to verify the accuracy of the algorithm and its effectiveness for cold start users.