Solving non-decomposable objectives using linear programming layers in general machine learning models : SVMs and deep neural networks

Abstract

Many domain specific machine learning tasks require more fine tuning with respect to nondecomposable metrics to be effective. In many applications such as medical diagnosis and fraud detection, traditional loss measures do not provide for the best performance. Optimizing for accuracy can give a false sense of effectiveness. In the case of Medical diagnosis, it is typical we want to minimize for false negatives. In the case of fraud detection, the data can often be imbalanced. In the case of imbalanced data, accuracy does not provide the best measure because the model may have high accuracy in the majority class while having low accuracy in the minority class causing the model to appear to have a high effectiveness on the data when in reality the bias of imbalanced data is skewing the results. For such cases, Non-Decomposable measures such a F-Score and AUC provide more detail into the real world performance of the model. Optimizing Non-Decomposable measures has posed a challenge in the past. In this paper, we will investigate using a proposed drag and drop linear programming layer for machine learning methods. We evaluate performance across multiple models and multiple data sets. Our results show increases in Non-Decomposable measures over traditional results and fast convergence.