Winsorised gini impurity: A resistant to outliers splitting metric for classification tree
Constructing a classification tree is sometimes complicated due to outliers occur in the data. Eliminating the outliers is the simplest option, but some important information will lose. Alternatively, one may make some amendments on the value of outliers, but the amended value is arguable in term of...
Saved in:
| Main Authors: | , |
|---|---|
| 格式: | Article |
| 出版: |
IP Publishing LLC
2014
|
| 主題: | |
| 在線閱讀: | http://repo.uum.edu.my/25772/ http://doi.org/10.1063/1.4903661 |
| 標簽: |
添加標簽
沒有標簽, 成為第一個標記此記錄!
|
| 機構: | Universiti Utara Malaysia |
| 總結: | Constructing a classification tree is sometimes complicated due to outliers occur in the data. Eliminating the outliers is the simplest option, but some important information will lose. Alternatively, one may make some amendments on the value of outliers, but the amended value is arguable in term of its suitability for classification purposes. We describe a strategy in order to identify and to handle the outliers in the process of constructing a classification tree. A Winsorised approach is suggested in estimating the impurity of the data prior to the splitting of each node of a tree. The proposed estimator provides a splitting value that resistant towards outliers in the data hence influences the performance based on plug in error rate of the tree. We examine the proposed idea on some real data sets represent various sizes of sample. The performance indicates that the proposed strategy is competitive, and sometimes shows better performance than traditional tree. |
|---|