Random Tag Forest

We consider the multi-label classification problem in this paper. We propose a randomized ensemble learning algorithm, random tag forest, which is an ensemble of random tag trees. Each tree is built by randomly defining a hierarchical tree structure over a subset of tag vocabulary. Each node in the tree corresponds to a tag in the vocabulary. During testing, a testing example will pass through each tree in the random tag forest. The tags along its path will be output as the prediction to this example. The final classification prediction is made by aggregating output across all trees in the random tag forest. The proposed approach is evaluated on a benchmark of image annotation. The experiments show that our approach achieves better results than the baseline.