Daniele The typical casue for that behavious is when you start assigng faces to a person with very low quality images - dark, low resolution, face is obscured or person is not looking into camera. This way program has difficulties to reliably extract face features and it may result into garbage suggestions. It is always recomended to start with good quality faces first, build a solid portfolio of faces that program will rely on when looking for similar ones, and only after that you can go on and confirm poor quality faces.
Hope that helps.
