Label Correction Strategy on Hierarchical Multi-Label Classification

One of the most popular approaches to solve hierarchical multi-label classification problem is to induce Support Vector Machine (SVM) for each class in the hierarchy independently and employ them in a top-down fashion. This approach always suffers from error propagation and yields such a poor performance of classifiers at the lower levels since no label correlation is considered during the construction. In this paper, we present a novel method called "label correction", which takes label correlation into consideration and corrects the results of unusual prediction patterns. In the experiment, our method does not only improve prediction accuracy on data in hierarchical domains, but it also contributes such a significant impact on data in multi-label domains.