Personalized Tag Predition Boosted by BaggTaming A Case Study of the Hatena Bookmark

We stated a learning problem, which we call taming, and develop a method for this problem in [神嶌 08b, 神 嶌 08c, Kamishima 08a]. The learner for this taming requests two types of training data sets, tame and wild. The labels of tame data is highly consistent with a target concept, which we actually want to learn. In contrast, wild data are not so well maintained; thus, some labels are consistent with the target concept, while some others are not. Additionally, we assume that wild data are much more abundant than tame data. This assumption is reasonable, because it is generally difficult to provide a large tame data set due to its high maintenance cost. The goal of the taming is to acquire more accurate classifiers by exploiting wild data. To achieve this goal, we developed a BaggTaming method, which is a modified version of bagging [Breiman 96]. We applied this BaggTaming to a personalized tag prediction for the data set collected from the delicious∗1, and showed the effectiveness of our method. We expect that this taming technique would be generally helpful for a tag prediction task on another collaborative tagging service. To check this hypothesis, we performed the tag prediction task on the hatena bookmark∗2, which is one of popular social bookmarks in Japan.