Font Size: a A A

Improvement Of Collaborative Filtering In Personalized Recommender Systems

Posted on:2021-09-01Degree:MasterType:Thesis
Country:ChinaCandidate:H Y CaiFull Text:PDF
GTID:2517306503491344Subject:Applied Statistics
Abstract/Summary:PDF Full Text Request
With the rapid development of the economy and the Internet in recent years,there has been a phenomenon called "data explosion",which results in information overload.As a result,a recommendation system has been proposed to solve the problem.This article mainly aims at improving the user-based collaborative filtering algorithm(user-CF).Firstly,the article summarizes some mature recommendation algorithms,as well as the advantages and disadvantages.Secondly,for the problem of user cold-start existing in user-CF,a similarity calculation model based on user profile has been established.Thirdly,to solve the problem of high-dimensional matrix when calculating users' similarities,a hybrid model has been established.The experiments are based on the Movie Lens-100 k dataset,using 5-fold cross-validation.The recommended form is scoring prediction,and the feasibility can be verified according to MAE and RMSE.In the experiments,the classic user-CF is firstly implemented,and the recommended result is movie IDs.The article explores the impact of different k on recommendations(k is the number of similar users of the aim user),and finds that 6)= 20 is what we want.So,the result at this time is regarded as the control group.Secondly,a similarity calculation model which is based on user profile is implemented,comparing the recommendation results under movie IDs and movie characteristics.Compared with the control group,the result of movie IDs is slightly worse,however,the result of movie characteristics is much better,which proves the feasibility of the new algorithm.Finally,a hybrid model of contentbased recommendation and user-CF has been implemented.Many similarity methods such as cosine,Pearson,Manhattan distance,Canberra distance,and Kendall have been tried.The accuracies of these different similarities are almost all better than the control group,which proves the feasibility of the new algorithm.In addition,by comparing the different recommendation results,the applicable scenarios of each similarity calculation formula can be summarized.
Keywords/Search Tags:collaborative filtering, user similarity calculation, user profile, User-item's main characteristic rating matrix
PDF Full Text Request
Related items