Improving the stability of variable selection for industrial datasets