Research On Recommendation Algorithm Based On Regression Strategy

The Internet provides extensive and complex information, and ordinary users can hardly capture, understand or use information effectively limited by the knowledge reserve and cognitive ability. Recommendation system based on big data mining technology can make user’s interests model through analysis of user’s behavior on the internet in the history, and actively recommend to user information satisfied to his demand without clear user requirement. Typically, the various forms of demand is defined as item, and interests is represented by rating, and system decide whether to recommend item to the user based on his rating level. How to predict unknown user ratings by the user history rating records, which called rating prediction problem, is the core of recommendation system research.Firstly, proposition and development of collaborative filtering, content-based and hybrid recommendation approaches are described in the paper, then, the important roles in recommendation system played by regression are pointed out.Secondly, the theory and development of traditional regression approaches such as linear regression and k-nearest neighbor regression are deeply discussed in the paper, as well as NCLUS network regression approaches.Thirdly, a new method for network regression is proposed, called IWR, is proposed in the paper. Weighted regression is taken as local predictor during an iterative learning process. The predicted labels are changed each step until meet the requirement. In addition, for the treatment of discrete attribute and method of attribute selection, it draws lessons from linear regression.Fourthly, this paper explores autocorrelation between items using rating dataset, and views items as nodes of network, and redefines the distance between nodes, then converts rating prediction problem into network regression problem.Finally, IWR is applied to rating prediction problem.Experiment is divided into two parts. First part with spatial and social networks shows that the proposed method is effective compared to traditional regression algorithm as well as NCLUS. Second part with MovieLens dataset shows that the proposed method is effective applied to rating prediction and performs better than Item-based neighborhood collaborative filtering.