Study On Trust Relationship Based Collaborative Filtering Recommender Strategy

As the flourish of E-commerce, information overload has been the most urgent issue for the network users. The birth of recommender system has become a prevalent way to deal with this problem; it provides users with personalized active service through analysis of users' preferences and relationship between items. Recommender systems are divided into two categories: content-based recommender system and collaborative filtering (CF) recommender system. The latter is the most common used technique in personalized recommendation, which can be also divided into user-based CF and item-based CF. Traditional collaborative filtering refers to user-based one.Through computation of user similarity from ratings on different items, user-based CF chooses the most similar users and forms the neighborhood of the active user. With reference of neighbors' ratings, recommendation list can be generated including items with high prediction. Traditional user-based CF has many drawbacks, such as data sparsity, cold start, which seriously influence the accuracy of recommendation. On the one hand, user similarity is difficult to calculate under the circumstance of data sparsity; on another hand, if common rating is too rare, similarity obtained is somewhat inaccurate, Further more, since cold start users contribute little to the system, relationship with other users is hard to mine, which leads to the trouble of neighborhood formation, as well as the discount of recommendation service..To deal with the problems we just mentioned above, a method introducing trust relationship into traditional CF recommendation process is proposed in this paper. Parting from traditional CF, user similarity is combined with user trust value and join together to produce a compound value, at which the recommendation is given. Further more, a set of trust propagation rules have been defined, using which trust relationship can be propagated through trust networks and more neighbors could be matched for cold start users. In this case, the problem that the system can not make recommendation because of short of ratings could be effectively alleviated.To evaluate the validity of the model, three experiments have been conducted, which aims at different purposes individually. Based on the dataset chosen in this paper, the experiments show that trust network based CF recommendation strategy get a better performance than traditional CF method in the case of data spasity., recommendation accuracy has been obviously improved.