Fig. 2.

Cluster metrics: graphical visualization. Blue stars represent samples predicted as binding by the model we want to explain. Red crosses are samples predicted as non-binding. The thick continuous line is the decision boundary of the model. We can identify three clusters for the binding samples, and two for the non-binding ones. As an example, we graphically visualize the three levels of metrics applied to the non-binding anchors. The black crosses denote the medoids of the (non-binding) clusters. Dotted circles denote the boundary decisions of the computed anchors: all the samples within the circle fulfill the anchor rule. Red rectangles highlight the true positives according to the anchor rules. CL metrics considers true positives the samples that fulfill the anchor rule and belong to the same cluster as the medoid used to compute the rule. CS level metrics considers true positives all the samples that fulfill the anchor rule and belong to the same split as the medoid considered. Finally, SL metrics considers true positives all the samples that fulfill any of the anchor rules of a split and belongs to that same split