Multi-label graph-induced error is computed by:

- Finding the smaller set between the true and the predicted classes of each document.
- Computing the minimum graph distance between each class of the smaller set and the closest class of the other set, in such a way that minimizes the sum of distances.
- Setting all classes in excess equal to the maximum distance.
- Adding all the distances and dividing them by the number of the classification tasks. The number of classification tasks is equal to the sum of the true categories of each document.

In our experiments we define the maximum distance as five, so all distances above five are treated the same.