On the sample complexity of semi-supervised multi-objective learning