Winsorize tree algorithm for handling outliers in classification problem
Classification and Regression Tree (CART) is designed to predict or classify the objects in the predetermined classes from a set of predictors. However, having outliers could affect the structures of CART, purity and predictive accuracy in classification. Some researchers opt to perform pre-pruning...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English |
| Published: |
2016
|
| Subjects: | |
| Online Access: | https://etd.uum.edu.my/5780/1/depositpermission_s92068.pdf https://etd.uum.edu.my/5780/14/s92068_01.pdf https://etd.uum.edu.my/5780/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Universiti Utara Malaysia |
| Language: | English English |
| Summary: | Classification and Regression Tree (CART) is designed to predict or classify the objects in the predetermined classes from a set of predictors. However, having outliers could affect the structures of CART, purity and predictive accuracy in classification. Some researchers opt to perform pre-pruning or post-pruning of the CART in handling the outliers. This study proposes a modified classification tree algorithm called Winsorize tree based on the distribution of classes in the training dataset. The Winsorize tree investigates all possible outliers from node to node before checking the potential splitting point to gain the node with the highest purity of the nodes. The
upper fence and lower fence of a boxplot are used to detect potential outliers whose values exceeding the tail of Q ± (1.5×Interquartile range). The identified outliers are neutralized using the Winsorize method whilst the Winsorize Gini index is then used to compute the divergences among probability distributions of the target predictor’s
values until stopping criteria are met. This study uses three stopping rules: node achieved the minimum 10% of total training set, |
|---|