Cost-sensitive classification with deep neural networks

Abstract

Traditional classification focuses on maximizing the accuracy of predictions. This approach works well if all types of errors have the same cost. Unfortunately, in many real-world applications, the misclassification costs can be different, where some errors may be much worse than others. In such cases, it is useful to consider the costs and build a classifier that minimizes the total cost of all predictions. Earlier, cost-sensitive learning has received very little research with balanced datasets. Mostly, it has been mostly considered as one of the measures that solves the class imbalance problem. As the basis of the class imbalance problem is similar to costsensitive learning, we can mainly rely on the research done regarding the class imbalance problem. The purpose of this thesis is to experiment on how successful different cost-sensitive techniques are at minimizing the total cost compared to an ordinary neural network. The used techniques involve making neural network cost-sensitive based on the output probabilities. Additionally, oversampling, undersampling and loss functions that consider the class weights are used. The experiments are performed on 3 datasets with different degrees of difficulty and they involve binary and multiclass classification tasks. Also, 3 different cost matrix types are considered. The results show that all the techniques reduce the total prediction cost compared to an ordinary neural network. The best results were achieved using oversampling and cost-sensitive output modifications for both binary and multiclass case.