Tag recommendation based on multi-threshold continuous condition random field
SU Yi-dan · Jisuanji yingyong yanjiu · 2013
As the quality of recommendation results by tag recommendation system was not high,it would influence and mislead users to search and locate their required resources.And even information confusion would exist.To enhance the accuracy and coverage of the results,this paper proposed multi-threshold continuous condition random fields model.The model not only maintained the advantages of condition random fields: dispensed with independence hypothesis for data,but could avoid the label bias problem.Meanwhile this work also employed the co-occurrence rate between tags,the semantic similarity of tag pairs,and the user similarity three thresholds to extract tag features.Here concurrently dug out the dominant and recessive tags,fully combined with user differences.Through the maximum likelihood estimation method iterative calculation to get model parameters,then the work established the model to recommend tags.Tests in BibSonomy data set show that this method is feasible.The result comparisons with the continuous condition random field model and the maximum entropy model display that tags recommended by this model are more accurate and more comprehensive.The stability of the model performs well.In the future,this work will be devoted to shortening the training time of the model.