Research And Implementation Of Movie Recommendation Algorithm

With the rapid development of Internet， the amount of information on the Internetgrows explosively. Users of the Internet are facing with the challenge of InformationOverload. As an effective approach to overcome this challenge， the booming personalizedrecommendation system is applied to various kinds of services on the Internet. Based on theidea of collective intelligence， the traditional collaborative filtering-basedrecommendation algorithm utilizes interaction information between users and items， suchas purchase， rating and so on， to measure the preference of users on items. In order toimprove the performance of recommendation system， the idea of collaborative filtering isintroduced to social network relation between users and improve the accuracy ofrecommendation by integrating social relation and rating relation.In personalized recommendation system， besides the interaction information betweenusers and items and the social relation information between users， the relative attributeinformation of users or items can be used to improve the performance of recommendationsystem. To integrate various type of relative information effectively is a new challenge forresearchers of personalized recommendation system. In this paper， we introduce theDouble-Star Graph to represent the complicated personalized recommendation system andpropose a novel matrix factorization-based method called Double-Star Graph-based matrixfactorization. Our proposal improve the performance of personalized recommendation byadding a similarity regularization term of the relative attribute information of users anditems， which aim to integrate effective relative information into matrix factorization.To evaluate the effectiveness of the novel approach， we crawled a dataset fromDouban Movie site. The dataset contains not only ratings between users and movies， butalso the social relation between users， user attributes (location and interest groups of users)and movie attributes (type， actors and directors). We also implement some start-of-artrecommender approaches， including probabilistic matrix factorization-based collaborativefiltering， social matrix factorization-based collaborative filtering and multi-relation factorization-based collaborative filtering. We perform experiments， includedeffectiveness experiments and cold-start experiment， on a Douban Movie dataset betweenour proposal and several start-of-art recommendation methods. The experiments resultsshow that our proposal outperform other competitor in rating prediction， as well as inalleviating the cold-start problem.