Multi-label Classification Using Random Label Subset Selections

摘要

In this work, we address the task of multi-label classification (MLC). There are two main groups of methods addressing the task of MLC: problem transformation and algorithm adaptation. Methods from the former group transform the dataset to simpler local problems and then use off-the-shelf methods to solve them. Methods from the latter group change and adapt existing methods to directly address this task and provide a global solution. There is no consensus on when to apply a given method (local or global) to a given dataset. In this work, we design a method that builds on the strengths of both groups of methods. We propose an ensemble method that constructs global predictive models on randomly selected subsets of labels. More specifically, we extend the random forests of predictive clustering trees (PCTs) to consider random output subspaces. We evaluate the proposed ensemble extension on 13 benchmark datasets. The results give parameter recommendations for the proposed method and show that the method yields models with competitive performance as compared to three competing methods.