Enhancing Accuracy of Multilabel Classification by Extracting Hierarchies

A novel algorithm of extracting hierarchies with the maximal F-measure for improving multilabel classification performance, the PHOCS, builds Predicted Hierarchy Of Classifiers. Nodes contain classifiers, and each intermediate node corresponds to a set of labels, and a leaf node to a single label. Any classifier in the extracted hierarchy deals with a considerably smaller set of labels as compared to the number L of labels, and with a more balanced training distribution. This leads to an improved classification performance. Our method has linear training and logarithmic testing complexity with respect to L. The experiment was conducted on 4 multilabel datasets and it has confirmed the effectiveness of the PHOCS algorithm.