Table 3.

Global statistics comparison between TBGA, BioRel [24], and DTI [10] datasets

Dataset	Split	Instances	Bags	Inst.s/bag	Relations
BioRel	Train	534,277	39,969	13.37	125
	Validation	114,506	20,675	5.54
	Test	114,565	20,756	5.52
DTI	Train	604,303	472,033	1.28	6
	Validation	6133	4769	1.29
	Test	6312	4817	1.31
TBGA	Train	178,264	85,047	2.10	4
	Validation	20,193	10,491	1.92
	Test	20,516	10,494	1.96

Statistics are reported separately for each data split. Columns represent, from left to right, the considered granularity level, the data split, the total number of instances and bags, the average number of instances per bag, as well as the total number of relations