Linking data separation, visual separation, and classifier performance using pseudo-labeling by contrastive learning