Prediction of Breast Cancer from Imbalance Respect Using Cluster-Based Undersampling Method

Area of Science:

• Medical Informatics
• Machine Learning
• Data Science

Background:

• Class imbalance is a significant challenge in breast cancer diagnosis, potentially leading to biased classification models.
• Existing machine learning methods struggle to accurately identify minority class instances in imbalanced datasets.

Purpose of the Study:

• To propose a novel hybrid classification method, K-Boosted C5.0, for addressing two-class imbalanced problems in breast cancer diagnosis.
• To enhance classification performance by leveraging undersampling techniques guided by K-means clustering.

Main Methods:

• A hybrid approach combining K-means clustering for informative sample selection and Boosted C5.0 as the classifier.
• K-means algorithm clusters majority and minority instances, selecting a balanced number of samples near class boundaries.
• Boosted C5.0 classifier is trained on the undersampled dataset to improve predictive accuracy.

Main Results:

• The K-Boosted C5.0 hybrid method demonstrated superior performance on 14 diverse datasets (12 small-scale, 2 large-scale).
• Evaluated using Matthews' correlation coefficient (MCC) and accuracy, the proposed method outperformed several competitive algorithms.
• Instance selection via clustering enhanced training data diversity, contributing to improved classification outcomes.

Conclusions:

• The K-Boosted C5.0 hybrid method offers a robust solution for imbalanced breast cancer datasets.
• This approach presents a promising alternative to conventional machine learning techniques for improved diagnostic accuracy.
• The method's effectiveness in handling class imbalance suggests broader applicability in similar biomedical classification tasks.