User Modeling in Folksonomies: Relational Clustering and Tag Weighting

This paper proposes a user-modeling method for folksonomic data. Since data mining of folksonomic data is difficult due to their complexity, significant amounts of preprocessing are usually required. To catch sketchy characteristics of such complex data, our method employs two steps: (1) using the infinite relational model (IRM) to perform relational clustering of a folksonomic data set, and (2) using tag-weighting to extract the characteristics of each user cluster. As an experimental evaluation, we applied our method to real-world data from one of the most popular social bookmarking services in Japan. Our user-modeling method successfully extracted semantically clustered user models, thus demonstrating that relational data analysis has promise for mining folksonomic data. In addition, we developed the user-model-based filtering algorithm (UMF), which evaluates the user models by their resource recommendations. The F-measure was higher than that of random recommendation, and the running time was much shorter than that of collaborative-filtering-based top-n recommendation.