The Research On User Modeling For Internet Personalized Services

With the explosive growth of information available on the Internet, there is a clear demand for personalized services, which provide users with information to their interests. As the core technology of personalized services, user modeling has recently attracted much attention.The rise of personalized services is described and three main forms of personalized services are introduced. The expression methods of user model are summarized and the information employed in user modeling is analyzed. The techniques of user modeling are induced. According to the users' engaging degree, the techniques of user modeling are categorized into three classes, i.e. user custom modeling, supervised modeling and automatic modeling.After analyzing the supervised user modeling, which is widely used currently, supervised user modeling employing users' backgrounds is proposed. Three rough interest-granularity supervised user modeling approaches based on information gain, rough set theory and genetic algorithms respectively are presented. Empirical results show that the three approaches perform well. Documents are described in information theory, based on which a fine interest-granularity supervised user modeling method is proposed. Experiments demonstrate that the approach has good performance. Since supervised user modeling employing users' backgrounds reduces the extra work from the user, it improves the performance of user model and makes personalized systems friendlier.Users' browsing behaviors can reflect users' interests. According to the characteristics of browsing behaviors, users' browsing behaviors are categorized into three classes, i.e. physiological behaviors, explicit behaviors and implicit behaviors. Implicit behaviors are the main source behaviors in the estimation of users' interesting degree. The Spearman rank correlation test and Kendall r rank correlation test are utilized to analyze the correlation of the typical implicit behaviors. The results show that the resident time on pages correlates with all other implicit behaviors. Consequently, the resident time on pages can approximately substitute all other implicit behaviors. Taking together the explicit behaviors into consideration, the minimal behaviors set for the estimation of interesting degree is obtained. Resident-time-based estimation and browsing-speed-based estimation are proposed. Empirical results show that the two approaches perform quite well and outperform the previously proposed method.After analyzing the current automatic user modeling approaches, automatic modeling based on user interests clustering is proposed. The characteristics of user interests automatic clustering are analyzed, which highlight that most of the classical clustering algorithms can not address the requirements of user interests automatic clustering. The ideas of the classical clustering algorithms are studied. According to the clustering idea, the classical clusteringalgorithms are categorized into two classes, i.e. cluster integer oriented clustering and object individual oriented clustering. A novel object individual oriented clustering algorithm, NEOREN, is proposed, which is based on the graph theory. Experiments demonstrate that NEOREN algorithm yields accurate results for clusters with arbitrary shapes, variable sizes, variable densities and outliers. The prominent advantage of NEOREN algorithm is that the data would be clustered into more clusters rather than be wrongly clustered even though the nearest neighbors number k doesn't reach the best value. The prominent advantage of NEOREN algorithm makes NEOREN fit for the applications sensitive to cluster errors rather than cluster numbers. The parameter k can be empirically set in the algorithm according to the application area; thereby automatic clustering can be achieved without requiring any extra work from the user. NEOREN algorithm is applied to user interests automatic clustering. Experimental results show that the user interests clustering approach based on NEOREN algorithm can discover the users'...