Study On The Algorithm Of Online Recommendation System And Its Application

In recent years, as the rapid development of the Internet, and personal digit devicessuch as smart phones and cameras quickly spread, large amounts of information isdigitalized and uploaded to the Internet. As the rapid increasing of Internet information,information overload also becomes a serious problem. People faced huge amounts ofdata every day, but they find it difficult to quickly obtain the information they reallyneed, and are often surrounded by various kinds of useless information. People urgentlyneed a way to actively, conveniently, accurately access and process information.The recommended system is one of the latest attempts to address this problem, andbecome a research hotspot in recent years. Such a system tries to actively or passivelycollect all kinds of user information, such as ratings of the items, or the web pages theuser read, the keywords the user searched, the links the user clicked, and try to infer theuser’s preferences. Then provide a list of recommend items to the user. However, theexisting recommendation systems still do not fully meet the needs of users in terms ofaccuracy, function and scope of application. New recommendation algorithms are stillemerging.Based on above background, we studied state-of-art recommender systems andtheir applications. This thesis focuses on the evaluation of recommendation system, andrecommender systems based on social network, and proposed some of improvedalgorithms.The work in this thesis includes the following aspects:1, The evaluation methods of recommender system are studied. The biases existingin current evaluating methods, and their impact on the overall system performance isanalyzed. Then an improved algorithm based on separating factors method to adjust thiskind of bias is proposed. The simulation experiment is carried out on the data of KDDCup2012Track2.2, The role of recommendation system in social network is studied. Then a trustpropagation and aggregation mechanism is proposed, to measure the correlationbetween two users in social network. An algorithm to quickly calculate this measure is also proposed. Simulation results of this algorithm on the KDD Cup2012Track1dataare also presented.