Analysis Of Recommendation Network And Research Of Personalized Recommendation Algorithms

This work studies the rating prediction problem of recommender system based on complex network theory. We mainly focus on the heterogeneity of users’ rating behaviors, and propose several efficient recommendation algorithms. Our work promotes the research of recommender system from personalized recommendation algorithms to personalization of the algorithm itself. The main work is as follows:1. We proposed a recommendation algorithm associated with user’s preference. The recently proposed hybrid algorithm assigned each user a uniform preference for algorithms, which went against the truth that users’ preference for algorithms are heterogeneous. This paper proposes and designs users’ personalized hybrid parameters which are associated with their historical behaviors. The results on two real data sets indicate that the new algorithm can get a higher 1.56% and 8.33% accuracy compared to the original one.2. We propose the concept of personalization of recommendation algorithms and study users’ best fit algorithm. Traditional algorithms apply single algorithm on all users, but the truth is that different users suit different recommendation mechanisms. This paper try to map users with their best fit algorithms based on users’ characters. Results indicate that if all users apply their best fit algorithm, it can get a higher 20.53% and 12.35% ranking accuracy compared with two typical algorithms. This work pushes for the solution of the relevance between data characters and best fit algorithms.3. We propose a recommendation algorithm based on item quality and user rating preference. The collaborative filtering algorithms face serious challenges of scalability when with large number of users or items. We propose a recommendation algorithm based on item’s quality and user’s rating preference, which has advantages such as low complexity and good interpretability. Results on three standard data sets show that the new algorithm will get an improvement of 2.99%, 3.18% and 3.53% in terms of prediction accuracy compared with the tendency-based algorithm.4. We propose a recommendation algorithm based on the reliability of users’ ratings. Since recommender system can directly affect users’ purchase intentions, so it’s vulnerable to be attacked by malicious users. In such case, it’s unreliable to get item quality by simply averaging the ratings. In this paper, we introduce several mechanisms based on user’s reputation to get the reliability of ratings, and propose a recommendation algorithm which can avoid malicious manipulations. Experiments show that the new algorithm can achieve 1.70% and 1.20% enhancement compared to the original one faced with strong attack.