node v1, and three to node v2. Table 5 shows which labels can be
assigned to which nodes and the accuracy of these assignments.
Note that the label distribution in node v2 is better than in v1. This
is due to the fact that when the attached labels of both nodes have
the same accuracy, the concept label attached to node v2 would be
more specific than that attached to v1. For example, if we attach labels
hComputeri to v1 and hLaptopi to v2, node v2’s label would
then be more specific than that of v1, even though they have the
same accuracy.
The example above indicates that the traditional entropy measure
is unable to measure the appropriateness of nodes when the
class labels are hierarchical. To overcome this weakness, this paper
proposes not only a new entropy measure suitable for measuring
the appropriateness of a node with hierarchical class labels, but
also a new strategy to determine the most suitable concept label
in the hierarchy for a leaf node.