Research On Personalized Recommendation Based On Random Walking In Collaborative Annotation System

Network Information has been getting unprecedented explosive growth with the rapid development of internet technology and the widespread popularity of smart terminal, which, however, brings a few problems. On the one hand, users can hardly get access to what they exactly need from the tremendous amount of information. On the other, plenty of valuable and useful information is covered up as a result of few eyes focus. Recommendation System that makes recommendation through matching the feature of user and information becomes one of the important technical means to solve these problems. Depending on different algorithm, Recommendation System can be divided into collaborative filtering recommendation, content-based recommendation, knowledge-based recommendation and hybrid recommendation. Nevertheless data sparseness resulted of huge information turns to be a soft spot of these traditional recommendation systems.Random Walk with Restart model is well acknowledged as an important means to figure out the problem of data sparseness. Tag, as an important link connecting resource and user in Folksonomy collaborative tagging system, can be used not only for classification and management of information but also for the feature description of users. Based on these two models, this paper put forward a personalized recommendation system based on bipartite graph with weight and resource cluster. It, taking a full consideration of different functions of User, Tag and Resource, make the recommendation through two step random walk. And the recommendation accuracy is greatly improved indicated by experiments on the dataset of Last. FM and MovieLensAccuracy has always been regarded as a main index to evaluate the recommendation result, meanwhile diversity becomes an important aspect when it comes to recommendation quality. This paper will also introduce an evaluation method under heterogeneous model, considering both accuracy and diversity, based on which it adopts greedy strategy and finds an approximate optimal diversified sub-set of the recommendation list. And this paper assembly this method with personalized recommendation model of this paper through the means of progressive strategy of the hybrid recommendation. The diversity of the recommendation results has been effectively improved suggested the experiment.