Abstract: Aiming at the problem that the classification accuracy and generalization ability of model were low in single variable selection algorithms, we proposed a hybrid variable selection algorithm. The algorithm was divided into two stages. In filtration stage, mutual information was used to quickly exclude a part of irrelevant variables, which reduced the dimension of sample space. In wrapper stage, the random forest was used to refine the remaining variables in the framework of permutation theory. The experimental results show that, compared with the contrast algorithm, this algorithm has higher classification accuracy and generalization ability.

Key words: hybrid algorithm, random forest, variable selection, mutual information