Balanced sensitivity and specificity on unbalanced data using support vector machine re-thresholding
Ridgway GR., Lehmann M., Rohrer J., Clarkson M., Warren J., Ourselin S., Fox N.
Support vector machine (SVM) classifiers use multivariate patterns to separate two groups by a hyperplane with maximal margin. This strategy tends to obtain good generalisation accuracy on even very high dimensional applications. However, SVMs are not well suited to unbalanced data with very different numbers of cases in each group. In this work we implement a properly cross-validated method for altering the SVM threshold (also known as the bias or cut-point) to re-balance the sensitivity and specificity.