Study On A Heterogeneous Objects Oriented Recommendation Method For Social Tagging Systems

Three different types of recommendation tasks could be seen in social tagging systems:tag recommendation, resource recommendation and user recommendation. These tasks have great value in helping user tag resources, gain information and enhance participation. However, problems of social tagging systems, such as the inconsistence in categorizing perspectives, the unguaranteed quality of tagging results, and the over sparse essence of tagging data, have make available recommendation methods unable to provide recommendations with high quality. Thus how to handle the aforementioned problems and provide recommendations of heterogeneous objects in social tagging systems has become a key research topic.This thesis studies a heterogeneous objects oriented social tagging recommendation method.In order to handle the inconsistence in categorizing perspectives and the unguaranteed quality of tagging results, this thesis studies preprocessing methods for social tagging data. By identifying classification and topical tags, and consensus and non-consensus tags, social tagging data are preprocessed and data quality is improved. To handle the over sparse problem, a unified recommendation method that combines the analysis of heterogeneous objects is proposed, and a corresponding personalized recommendation method. By introducing different types of objects with dense relations, more recommendation clues are leaded in.In details, considering the inconsistence in categorizing perspectives, this thesis analyzes and filters tag relations identified using tag semantic cases, and construct tag hierarchies to recognize classification and topical tags. For consensus and non-consensus tags, this thesis models tag semantics using keywords. A KeyGraph based text chance discovery method is used to tell consensus tags from non-consensus tags. To handle the over sparse problem. This thesis considers different types of objects with dense relations to introduce more recommendation clues and improves recommendation quality. Based on the LDA topic model, this thesis proposes a probabilistic generative model of tags, users, resources and contents in social tagging systems. Relations between heterogeneous objects are modeled using conditional probabilities. A parameter estimation method and a model inference method is then introduced to realize unified recommendation for heterogeneous objects. Based on such a model, a personalized recommendation method is then proposed. By extending the social tagging system probabilistic model, the preferences that a user uses different tags on resources is modeled to provide personalized recommendation with high quality.