A Recommendation System Combining Similarity Measurement And Feedback Adjustment

Along with the development of information technology and the popularity of the Internet, the public has imperceptibly stepped into the era of Information Overload. To deal with information overload, recommendation systems have attracted wide attention as they can provide the users with better personalized services. They can help the users find new and interesting information even when the users are not clear about whar they are looking for.A recommendation algorithm is the kernel of recommendation system. Although collaborative filtering is widely used, social network-based recommendation algorithms have become a new research topic along with research efforts in online social networks. In fact, people always turn to friends they trust for product recommendations, and their likes and dislikes are similar and can be easily affected by the friend circle they keep. Researchers in this area seek to introduce trust relationships from social networks into recommendation systems to simulate recommendations in reality.The contributions of this thesis are as follows.(1)This thesis review related work in recommendation systems, with a focus on collaborative filtering and social network-based algorithms, and analyze problems in existing research efforts.(2) This thesis discuss challenges in traditional similarity metrics in where there are not sufficient historical data for measuring similarity between products or users. To tackle this problem, we design a user similarity measurement that computes the similarity with a more comprehensive and objective point of view.(3) Suffering from the sparsity of trust relationship data, social network-based methods have to consider indirect neighbors that are only weakly trusted, and this affects their precision of recommendations due to long-tail noise disruption brought up by dissimilar users, In order to deal with this problem, we design a neighbor filtering method to select more similar neighbors.(4) The existing approaches all assume that review ratings are true and objective, and ignore that anomalous ratings also exist. Aimed at handling this problem, we introduce a feedback adjustment mechanism to detect and adjust anomalous ratings in order to increase the recommendation precision.(5) This thesis design and implement the recommendation system that combines similarity measurement and feedback adjustment. The system integrates the above methods, and it empirically demonstrated that it not only reduces the data sparsity problem and the dissimilar-user problem, but also increases the recommendation precision.