Interaction effect between data discretization and data resampling for class-imbalanced medical datasets.

Area of Science:

• Data Mining
• Machine Learning
• Medical Informatics

Background:

• Data discretization transforms continuous features to discrete ones, aiding specific data mining algorithms.
• Class-imbalanced medical datasets pose challenges for accurate classification.
• Data resampling techniques (oversampling, undersampling, hybrid) are used to balance training data.

Purpose of the Study:

• To evaluate the impact of combining data discretization and resampling on classifier performance for imbalanced medical datasets.
• To compare the order of applying discretization and resampling steps.
• To identify optimal preprocessing strategies for improved classification accuracy.

Main Methods:

• Experiments conducted on 11 two-class and 3 multiclass imbalanced medical datasets.
• Discretization algorithms: ChiMerge and Minimum Description Length Principle (MDLP).
• Resampling algorithms: Tomek links undersampling, Synthetic Minority Oversampling Technique (SMOTE), and SMOTE-Tomek.
• Classifiers: Support Vector Machine (SVM), C4.5 decision tree, and Random Forest (RF).

Main Results:

• Combined approaches yielded higher Area Under the ROC Curve (AUC) rates compared to baseline methods (0.8%-3.5% for two-class, 0.9%-2.5% for multiclass).
• For two-class data, MDLP discretization followed by SMOTE oversampling achieved the highest AUC with minimal computational cost.
• For multiclass data, SMOTE or SMOTE-Tomek resampling before ChiMerge discretization provided the best performance.

Conclusions:

• Oversampling techniques generally improve classifier performance over baseline methods.
• Data discretization alone does not guarantee improved classifier performance.
• Combining discretization and resampling offers potential for higher AUC rates on imbalanced medical datasets.