Variable Selection in Random Forest with Application to Quantitative Structure-Activity Relationship

A wrapper variable selection procedure is proposed for use with learning machines that generate a measure of variable importance, such as Random Forest. The procedure is based on iteratively removing low-ranking variables and assessing the learning machine performance by cross-validation. The procedure is implemented for Random Forest on some QSAR modeling examples from drug discovery and development. It is shown that the non-recursive version of the procedure outperforms the recursive version, and that the default Random Forest mtry function is usually adequate. The paper concludes with some comments about performance assessment and the dangers of using Random Forest’s outof-bag error estimate in a variable selection wrapper.