Tony OK, this is how it works. There are two stages - faces recognition and similar faces suggestion. Recognition phase is done once when image is scanned. It calculates features of the faces that are then used on the next stage.
The next stage is calculating new suggestions for the person. Everytime programs gets all current confirmed faces of that person, all rejected faces of same person, as well as all confirmed faces that do not belong to this person and using this information it calculates what faces could be good candidates to suggest for this person.
With that said, every time you confirm correct or incorrect face for the person, this affects the results that will be suggested next for that person. The less errors in the confirmed faces, the better suggestions will be. However, there is a limitation to how accurate suggestions could be, as the recognition is not perfect still. I will still struggle with dark or obscured faces, regardless what you do.
Hope that makes sense.