Research Of Video Recommendation Algorithm Based On Trust And User Behavior

With the booming development of the Internet, people now produce and have access to a large amount of data every day. It leads to low utilization of information, which is called the information overload problem. Recommendation system is an intelligent agent system to solve this problem. It recommends items to users based on their personal interests or needs, and provides fully personalized decision making and services. The core of recommendation systems is recommendation algorithm. After intensive study of recommendation algorithms, this thesis proposes a data cleaning approach to solve the scale problem of collaborative filtering algorithms, and a dependence-based social recommendation algorithm.Collaborative filtering algorithm’s scalability problem points to the situation, that with the increase of users’ data, the response time of the algorithm increases significantly. Data cleaning algorithm aims at removing items which do not represent the user’s interests so that user interests are extracted from a smaller but more efficient dataset. For this purpose, our approach improves the similarity’s calculation formula by considering the influences of the time period, the sequence of behaviors, and the asymmetry property among items. On this basis, our approach redefines a user’s interest values of the items in his history records, then clean up the items which have the lowest interest values in accordance with a certain set of rules, as they maybe interference data.This thesis defines an indicator of dependence degree among users in social recommendation algorithm, which consists of the trust value between two users, the target user’s credibility value and interest similarity value. The algorithm recommends items with the highest dependence value to users. To measure the interest similarity value, this thesis proposes a user-negative samples sampling method, and with this method our approach not only considers the contribution of positive samples, but also takes into account the contribution of negative samples. The trust value depends on users’ familiarity in the social network and their interest similarity values, and the credibility value depends on the mean of all other users’ trust values to him. The algorithm makes full use of the user’s social relationships and behavior information, combining the advantages of collaborative filtering recommendation and social network based recommendation. According to experimental results, its better performance was proved.To demonstrate the result of our work, a personalized video recommendation system was developed. And we successfully deployed the dependence based social recommendation algorithm of this thesis into the system. The system includes several modules such as user management, video management, data management and recommendation algorithm.