关键词: 不平衡数据集, 特征选择, 集成学习, 抽样

Abstract:

The traditional feature selection methods are basically aimed for getting the optimal accuracy without full consideration of the data distribution, which can not achieve promising results on imbalanced datasets. A new feature selection method was proposed based on the data distribution modification  for imbalanced data sets. This approach could modify data distribution  many times by sampling with replacement. The instances of large classes were equal to the minor class samples in each new dataset. Finally, the final selected features were generated by voting mechanism for ensemble learning, which could combine the selected features by receiving more votes   than half from all the new training datasets. Experimental results on several UCI datasets showed that the proposed method was an effective feature selection approach for imbalance problems.
 

Key words: imbalanced data, feature selection, ensemble learning, sampling