Improving reproducibility by controlling random seed stability in machine learning based estimation via bagging