Cost-sensitive prediction of airline delays using machine learning

This study provides a framework combining the sampling method called costing and supervised machine teaming algorithms to predict individual flight delays. The costing method converts cost-insensitive classifiers to cost-sensitive ones by subsampling examples from the original training dataset according to their misclassification costs. A weighted error function has been newly defined to evaluate the model's performance considering misclassification costs. And the function is measured by the various cost ratio between false positive error and false negative error. The cost ratio shows the relative importance of delays class to on-time class. The weighted error rate varies with the cost ratio and the model can have lower weighted error rate when the cost ratio is 10.