Transfer Labels

[JJ]

If we do a histogram on transfer labels, we can find that some
datasets have overlaps on different classes, some are not.
AVICENNA, 5 classes in transfer labels. A lot of overlaps (by overlap,
I mean a record is labeled by different classes as the same time)
HARRY: 4 classes in transfer labels. A small portion of class overlaps
RITA: 4 classes in transfer labels, no overlap between classes
SYLVESTER: 2 classes in transfer labels, no overlap
TERRY: 4 classes in transfer labels, some overlaps.

How do we explain the label overlaps in the transfer labels? Does
this happen in final and valid dataset too?

[Causality Workbench]

Some of the tasks are "multilabel" which means that a given pattern
can belong to several classes. This is quite common for instance in
text categorization (TERRY). For example if a document talks about a
fund raiser using a sport event to raise money for a political event,
it could belong to the categories "sports", "finances", and
"politics". For AVICENNA, some scripts could belong to several
classes. I do not know Arabic, but the Latin alphabet has that too:
"I" (Capital I), "1" (digit 1), and "l" (lowercase L) can all be
written as a vertical bar. For HARRY, the labels are actions in video
clips. Clearly several actions can occur in the same video (e.g.
someone opening a car door and someone running; someone hugging
someone else and someone drinking, etc.)

The organizers