Optimization of Area Under ROC Curve (AUC) via Ensemble Mini-Batch Approach

Optimizing the area under the receiver operating characteristic (ROC) curve (AUC) has garnered increasing attention in modern signal processing and machine learning applications due to the recent proliferation of high-volume and highly skewed datasets. Existing scalable approaches utilize linear models with limited modeling ability, kernel methods with high computational cost, or neural network models, which may be inadequate for the problem at hand. In this correspondence, we extend the mini-batch AUC optimization (MBA) algorithm called MBA to nonlinear settings and propose a new approach named ensemble MBA (EMBA). In this approach, we utilize MBA to learn randomized neural networks as base rankers, each of which entails a convex optimization problem. We then combine these base rankers in a principled manner via constrained logistic regression to learn an ensemble ranker/classifier. The performance benefits of the new approach are illustrated on real-world datasets.