Improved shrunken centroid classifiers for high-dimensional class-imbalanced data.

Area of Science:

• Machine Learning
• Statistical Classification
• Bioinformatics

Background:

• Nearest shrunken centroid (NSC) methods, including PAM, ALP, and AHP, are popular for high-dimensional data classification.
• Current NSC methods estimate shrinkage by minimizing overall cross-validated (CV) error rates.
• This approach can lead to biased classification, particularly with imbalanced datasets.

Purpose of the Study:

• To address the bias of NSC classifiers towards majority classes in imbalanced datasets.
• To propose a novel method for estimating shrinkage in NSC algorithms to improve performance on imbalanced data.
• To enhance the reliability and accuracy of classification for high-dimensional, imbalanced data.

Main Methods:

• Proposed estimating the amount of shrinkage by maximizing the CV geometric mean of class-specific predictive accuracies (g-means).
• Evaluated the proposed method on simulated and real high-dimensional class-imbalanced datasets.
• Compared the performance of the g-means approach against the traditional CV error minimization strategy.

Main Results:

• NSC classifiers exhibit bias towards the majority class when data are imbalanced.
• This bias is exacerbated by larger numbers of variables, greater class imbalance, and smaller inter-class differences.
• The proposed g-means approach significantly reduces bias and improves classification accuracy compared to minimizing overall error.

Conclusions:

• The g-means strategy effectively mitigates the class-imbalance problem in NSC classifiers.
• The proposed approach leads to superior performance over traditional methods for biased NSC classifiers.
• Utilizing the g-means approach results in a substantial reduction in the number of selected variables for NSC classifiers.