Research On Personalized Recommendation Algorithm Combining User Attributes And User-centric Natural Nearest Neighbor

At modern internet age,recommender system or recommendation engineis an effective way to address the information overload problem and recommend products or information to users,The basic underlying principle is to recommend products or information that meet users' demand or interests efficiently by extracting users' interests and preferences from analyzing the data consisting of users' historical behaviors and choices.Recommendation systems have been widely used in many fields such as Amazon product recommendation,Pandora music recommendation,Netflix movie recommendation,Google Reader recommendation,Facebook friend recommendation,sina news recommendation and many other areas.Currently,the main recommendation methods include collaborative filtering recommendation,content-based recommendation and hybrid recommendation.There are shortcomings in aforementioned methods,for instance,collaborative recommendation has the problem of cold start,and information of users and products is difficult to extract in method of content-based recommendation.A common solution is to combine more than two recommendation algorithms into a new hybrid recommendation method to overcometheir individual drawbacks.This paper first introduced the research background,reviewed the domestic and international research progress,summarized the development,current application,research focuses and existing problems ofpersonalized recommendation,and then compared the main recommendation methods as well as clustering methods used in recommendation.Building on this,a user attributes and user-centric natural nearest neighbor-based personalized recommendation algorithm(UA3NR)was proposed.The main achievements were as follows:1)As traditional collaborative filteringrecommendation based on K nearest neighbor(K-CF)doesn't consider mutual neighbor relationship,and parameter K is hard to identify,this paper proposed user-centric natural nearest neighbor recommendation(3NR)algorithm which self-adaptively searches target user's natural nearest neighbor set without the parameter K,and then recommend products to users with higher accuracy.While 3NR algorithmwas used to recommend products or information to target users,the concept active neighborswho play an important role inpredicting ratings of target user was proposed.The active neighbor set was added to the target user's natural nearest neighbor set to effectively alleviate the problem of data sparsity.2)Another user-clustering-based recommendation algorithm(UCR)was proposed to alleviate the cold start problem of 3NR.By clustering users by their attributes,UCR algorithm first finds out which cluster the target user belongs to and identifies the target user's neighbor set,and then predicts the target user's ratings and recommends products to the target user.UCR algorithm searches user's neighbors locally to lower algorithm complexity and raise recommend efficiency.In addition,by using user's attributes to find its neighbors,UCR can deal with cold start problem to a certain extent.3)3NR and UCR algorithmswere combined using the method of weighting aggregation into theUA3 NR algorithm,which ensures high accuracy,high efficiency and the ability to deal with data sparsity problem or cold start problem of new users.4)A series of experiments were conducted to compare the 3NR andUA3 NR algorithms with existing algorithms on MovieLens real datasets.Experiments show that the RMSE and MAE of 3NR is lower than that of K-CF and INS-CF algorithms,and the MAE of UA3 NR algorithm is lower than that of NPSSC algorithm on 5 to 15 clusters.