Recommender System Research Based On Newly Added Users And Ratings

The rapid development of computers and Internet has changed the way people live in.In the past,people communicated in reality.Nowadays,people have social activities on the Internet.The popularity and application of the Internet has made the Internet record a large amount of data.The huge amount of data makes it difficult for people to quickly select the information they are interested in.This is the problem of “information overload”.As an information filtering tool,the recommender systems are widely used because it provides users with accurate and personalized recommendations.Today,the number of users using the recommender systems is increasing.If new users without any historical behavior data enter the recommender systems,how to make them get recommendations and how to predict the ratings for them?In the network-based recommender systems of considering time information,there exists a problem that new users cannot acquire recommendations.In this paper,we investigate the effect of adding new user nodes to the network on the recommendation performance.Further,by applying the above conclusion to fill ratings for new users in the rating matrix,we study the impact of filling rating on rating prediction for the new users.The main work of this thesis is as follows:In the network-based recommender systems of considering time information,some users in the test set do not exist in the training set,and in this paper,we take this part of users as new users.In order to make new users get recommendations,we propose to add virtual new user nodes in the user-item bipartite network,and add links for the new users.In order to add links for the new users,we propose three schemes: the first scheme is to connect with the items the small-degree user links to in the training set,and the user is randomly selected from all the users whose degree is from 1 to 20.The second scheme is to connect with the items the median-degree user links to.The user is randomly selected from all the users whose degree is from the average degree minus 2 to the average degree plus 2.The third scheme is to connect with the items the large-degree user links to.The user is randomly selected from the largest 50 degree users.Based on two benchmark datasets,MovieLens1 M and MovieLens100 K,by employing six network-based recommendation algorithms,including CN,AA,Salton,Sorensen,MD,HHM algorithms,the effects of the three link-adding schemes on recommendation performance are compared.The results demonstrate the first scheme makes recommendation accuracy,diversity,and novelty better.(2)Further,adopting the first scheme to fill the ratings for the new users in the training set,namely,filling the actual ratings of items the inactive users are interested in into the training set,we investigate the rating prediction problem for the new users.By applying the demographic collaborative Filtering(DCF),SOREC methods,we find that the newly-added-rating SOREC algorithm(RSOREC),the newly-added-rating collaborative filtering algorithm(RDCF)and the newly-added-rating user-based hybrid collaborative filtering algorithm(RUHCF)has a higher prediction accuracy.In conclusion,adding new user nodes to the network-based recommender system of considering the time information,and connecting the new users with the items the small-degree user links to,it makes the recommendation accuracy,diversity and novelty better.Filling the actual rating of items which the inactive user has rated in the training set,it makes the rating prediction accuracy higher.