Utilizing Hierarchies in Tree-Based Online Structured Output Prediction

Methods for online prediction of structured values are becoming more and more popular. However, hierarchical prediction, which has recently been shown to produce good results in terms of predictive performance in the batch learning setting, has not yet been applied in the online learning setting. We address the recently introduced task of hierarchical multi-target regression. To this end, we propose a hierarchical extension of iSOUP-Tree, which can address online multi-target regression. The extension weighs the split evaluation heuristic according to the location of the targets in the hierarchy. We design the experimental setup to ascertain whether the additional information contained in the hierarchy can be utilized to improve the predictive performance in the leaf targets. The proposed method shows promising results, producing potential improvements that should be investigated further.