Research And Implementation For Collaborative Filtering Optimization Algorithm

With the development of Internet and E-commerce, the recommender system has gradually become an important research field of E-commerce technology, and attracts many researchers'attention. Collaborative filtering technology is one of earliest applied and most successful technologies in recommender system. And it is the main research issues in the field of personalized recommendation, and in this paper we focus our research on it.In this thesis, it is pointed out that with the development of E-commerce, the magnitudes of users and commodities grow rapidly, which resulted in the extreme sparsity of user rating data by analyzing the problems that exist in today's collaborative filtering technology. For emphasizing the impact of item's genres on similarity computing, the traditional similarity measure methods work inaccurately in this situation. For this problem, an idea of making item's genre information take part in similarity computing is proposed. It is applied in item-based and user-based collaborative filtering algorithms respectively. The former uses the matrix of item and item's genre for genre part, the latter uses the matrix of user and item's genre gained from the matrix of user rating and the matrix of item and item' genre for that part, then it is combined with each corresponding original similarity via linear approach to become similarity between items and between users:Experiments show it increases accuracy of prediction at different levels both in item-based and user-based collaborative filtering.And then a modified user-based collaborative filtering algorithm is proposed for solving the problem that the traditional collaborative filtering can't reflect the difference of user' attention to different kinds of items. The algorithm combines item-based and user-based collaborative filtering according to combined recommendation.This algorithm regards user-based collaborative filtering as the main body, using neighbors of item to be predicted produced from item-based collaborative filtering to select nearest neighbors of active user got from user-based collaborative filtering again. It is able to take the difference of user's interest in different kinds of items and find'real'neighbors of active users regarding every kind of items. Experiments show it can effectively avoid the shortcomings of traditional methods and improve the accuracy of prediction, thereby enhancing the collaborative filtering system on the recommendation quality.