Optimizing between data transformation and parametric weighting for stable binary classification

In this paper, an optimization problem is formulated for stable binary classification. Essentially, the objective function seeks to optimize a full data transformation matrix along with the learning of a linear parametric model. The data transformation matrix and the weight parameter vector are alternatingly optimized based on the area above the receiver operating characteristic curve criterion. The proposed method improves the existing means via an optimal data transformation rather than that based on the diagonal, random and ad-hoc settings. This optimal transformation stretches beyond the fixed settings of known optimization methods. Extensive experiments using 34 binary classification data sets show that the proposed method can be more stable than competing classifiers. Specifically, the proposed method shows robustness to imbalanced and small training data sizes in terms of classification accuracy with statistical evidence.