Abstract: Aiming at the problem of single feature evaluation criteria in feature selection process, we proposed a new feature selection algorithm based on the extreme gradient boosting algorithm in ensemble learning. Firstly, the metrics of building ensemble tree model in the extreme gradient boosting algorithm were used as the importance measures of features in feature selection, and then a new bidirectional search strategy was used to balance the influence of multiple feature importance on the results, and optimize the efficiency of evaluation process. Through the test of 11 different dimensions of standard datasets, the experimental results show that the algorithm can increase the diversity of feature subsets, accelerate the speed of feature selection, and has high computational efficiency on both medium and low dimensional datasets, and can deal with high-dimensional datasets.

Key words: feature selection, extreme gradient boosting, bidirectional search