*5.1.11 Hierarchy of multilabel classifiers (HOMER)*

HOMER [25] is an effective and computationally efficient for multi-label classification problem. Its principle consists of constructing a hierarchy of multi-label classifiers in the form of tree, following the divide-and-conquer strategy. Each deals with a much smaller set of labels compared with the set L. Each node of the tree contains a set of labels and the produced classifier, in which the root contains the set of all labels, and the leaves contain a single label. Each internal node contains the conjunction of the label sets of its children. The construction of this tree is done by following these steps:

1.The root contains all labels.

	- Create k children.
	- Each child filters the training dataset of its parents by keeping instances that have at least one of its own labels.
	- Train the classifier Hn using the filtered dataset.

The question is how to distribute the labels of a node on k children?

For each child, the labels may be evenly distributed to k subsets in a way such that labels belonging to the same subset are as similar as possible. To do this, HOMER uses a balanced clustering algorithm, known as balanced k means.
