Study On Tag Recommendation Methods For Social Tagging Systems

Web2.0has greatly simplified the way people share information over the Internet. Such simplicity gives more information a chance to get spread. But meanwhile, it also brings challenges to the organizing of information. Social tagging systems categorize information using a plain text label based approach. The high simplicity and usability of such systems have made them widely accepted, and have turned social tagging to be the most important information organizing method at the age of Web2.0. However, due to the uncontrolled nature of social tagging, the categorizing results of social tagging are still facing problems, such as the inconsistence in categorizing perspective, terms and results, the redundancy and incompleteness of categorizing results, and the lack of standardization of category usage. In order to enhance the quality of social tagging, tag recommendation, as a social tagging assisting method, has received lots of attention from the field of social tagging research.Lots of works have been done to solve the tag recommendation problem. Many approaches have been proposed by researchers, and good experiment results have been yielded with real life datasets. However, solutions for some key problems of tag recommendation are still lack. Firstly, available methods may either ignore the semantic information of tags, or rely on external semantic sources to model the semantics of tags. Due to the domain coverage, definition perspective and update cycle problems, external semantic resources are not suitable for large scale social tagging applications. Secondly, these methods lack a proper preprocessing of social tagging data, which leads to a bad data foundation for recommendation algorithms. Finally, these methods could not properly leverage various kinds of recommendation clues and could not model users’ personal preferences. These shortages limit the quality of recommendation results provided by these methods.To handle these disadvantages and to provide high quality tag recommendations, this thesis studies tag recommendation methods for social tagging systems. By leveraging detailed semantics of unpopular tags and the co-annotation relation between tags, a tag semantic model is proposed to provide solid semantic foundation for tag recommendation. By identifying classification and categorization tags, consensus and non-consensus tags, and extending resources’tags using relations, social tagging data preprocessing is applied to provide data foundation for tag recommendation. Based on these works, a hybrid tag recommendation method which analyzes multiple types of objects is proposed to solve the over sparse problem and the lack of recommendation clue problem in social tagging systems. An autonomy oriented personalized recommendation method is then introduced to give more personalized recommendations. In detail:(1) A tag semantic model based on semantic co-annotation. Based on the clear yet detailed semantic information of unpopular tags, tags’semantics are modeled as the semantics propagating according to the mutual annotation relation between tags. A formal math model and a calculation method are introduced, and the performance of the model, together with the parameters’impacts on the model, is evaluated by experiments. Such a model provides an effective way to capture the semantics of tags.(2) Preprocessing methods for social tagging data. A classification and categorization tag identification method based on tag semantic hierarchy recognition is proposed to handle the phenomenon of different topic coverage of tags. A consensus and non-consensus tag identification method is introduced to reflect the situation of consensus agreements of tags reached by users. A relation based resource tag expanding method is presented to deal with the non-uniform distribution of tags and the fact that lots of resources are lack of tag annotations. Each method is empirically evaluated by experiments, and results show that these methods could help giving sound data foundation for recommendation algorithms.(3) A tag recommendation method based on a combined analysis of heterogeneous objects. To resolve the over sparse problem and the lack of recommendation clue problem, a tag recommendation method that combines the analysis of heterogeneous objects is proposed. By introducing different types of objects with dense relations, more recommendation clues are leaded in. Details of the probabilistic system model, and the parameter evaluating, model inferring and incremental updating methods are described. Experiments analyze the performance of the method, and proves that compared with the other evaluated methods, the proposed method could effectively make use of more recommendation clues, and provide recommendation results with higher quality.(4) An autonomy oriented personalized tag recommendation method. To tackle the lack of modeling for users’ preferences on resource and tag usage, an autonomy oriented personalized tag recommendation method is introduced. By directly modeling to what extend the user is interested in a resource, and the preferences that a user uses different tags on resources, high quality personalized recommendation is achieved. Experiment results show that the autonomy oriented method could provide more personalized recommendation results in compare with other evaluated personalized tag recommendation methods.