Measuring algorithmic bias in face analysis — towards an experimental approach