allow negative samples (and regulate their amount for partial data)

[sfluegel05]

In partial data (e.g. ChEBIOver50Partial) a top class is selected and only subclasses of this top class are used as labels.
The samples are filtered s.t. only samples with at least 1 positive label (i.e., subclasses of the top class) are used as samples.

The problem

no generalisation beyond beyond the top class (I trained a model on 30 labels (subclasses of 22712) and got ~70% macro-F1 with only samples that are subclasses of 22712, but 10% macro-F1 when looking at samples from the whole ChEBI.

Solution

Remove filters that only allow samples with at least 1 positive label, add a ratio parameter external_data_ratio that determines the amount of negative samples (0 for n negative samples, the current behaviour, 1 for all possible negative samples)